AI4S · Simulation

The Landscape of Scientific Simulation: sim2real Gaps and AI

Simulators across 13 disciplines, from DFT and GraphCast to virtual cells and Lean: how big their sim2real gap really is, how to measure it, how they feed AI, and where the money flows.

Spark Science ResearchUpdated Sep 21, 202651 min read

From DFT, molecular dynamics and GraphCast to virtual cells, ESM3 and Lean: representative simulators across 13 disciplines, how large their sim2real gap is, which metrics should measure it, and how each relates to AI training.

The core regularity: how fast AI advances in a field is proportional to how verifiable its reward is, and inversely proportional to its sim2real gap.

13
Disciplines / directions covered
3
Generations of simulation · 3 layers of gap
5
Ways simulation relates to AI
~$50B
Engineering-simulation M&A, 2024–26

Figures marked ⓚ come from prior knowledge and were not re-verified for this edition.

Key takeaways

  1. "Simulation" actually comes in three generations, and they cannot be compared on the same footing.

    • Generation 1: first-principles numerical simulation. DFT, molecular dynamics, CFD, numerical weather prediction (NWP), Geant4, MuJoCo. The governing equations are known; error comes from approximations and limited compute.
    • Generation 2: AI emulators / surrogates. For example MLIPs, GraphCast and BioEmu. They are trained on Generation 1 output and run 10³–10⁵ times faster, but their accuracy ceiling is the teacher (the Generation 1 simulator).
    • Generation 3: data-driven "world models". Virtual cells, ESM3, Evo 2, the mouse visual cortex digital twin. These fields have no usable first-principles equations, so models learn directly from experimental data. ESM3 and virtual cells are Generation 3, and their gap looks completely different from DFT's.
  2. The sim2real gap has to be measured in three layers: (1) simulator vs. real experiment; (2) AI emulator vs. simulator; (3) the wet-lab success rate of whatever gets designed. The most common exaggeration is passing off (2) as (1). AI weather models score 15–45% lower error against ERA5 than against real weather stations; FEP has an RMSE of 1.73 kcal/mol on public benchmarks, which becomes 2.44 on pharma companies' private systems.

  3. Ranked from smallest to largest gap: mathematics (Lean; the verification-layer gap is ≈0, though kernel implementations have had bugs, see §2.13) < rigid-body robot control and plasma control (already deployed zero-shot on real hardware) < weather and fluids (the simulations themselves are already accurate) < materials and molecular energies (DFT itself carries systematic error) < protein design (wet-lab success 5–50%) < drug binding free energy < virtual cells (in 2025 they still did not reliably beat the naive "take the mean" baseline).

  4. Simulation already relates to AI in four forms, in descending order of maturity: simulation generates data to train AI (most mature) → AI replaces simulation → simulation as an RL environment for training control policies (the TCV tokamak, the type 1 diabetes artificial pancreas, robotics) → simulators or predictors as tools and rewards for LLM agents (ether0, Lean, virtual-cell agents). Frontier companies (Periodic Labs, Lila) are betting on a fifth form: "use nature itself as the RL environment." The reason is exactly point 3: most scientific simulators have too large a gap to serve as the ultimate reward.

Gen-2 AI emulator MLIP · GraphCast · BioEmu Gen-1 simulator DFT · MD · CFD · NWP Reality experiment · observation L2 L1 L3: does the designed thing work in reality? A beautiful L2 number does not tell you anything about L1
The three layers of sim2real gap. Gen-2 error stacks on top of Gen-1 error; only L1 and L3 touch reality.

The landscape at a glance

DisciplineFirst-principles simulation (Gen 1)AI emulator / foundation model (Gen 2/3)What is simulatedSource of ground truth
Cell biologyWhole-cell ODE model (Karr Mycoplasma, 2012 ⓚ), VCell ⓚArc State, scGPT/Geneformer; proteomics virtual cell (Nature 2026)Perturbation → transcriptome / cell statePerturb-seq single-cell sequencing
Protein / RNA / genomeRosetta (energy function) ⓚESM3, AlphaFold3, Boltz, RFdiffusion, BindCraft, Evo 2Sequence ↔ structure ↔ functionCrystallography / cryo-EM, wet-lab activity
Molecular dynamics / drugsGROMACS, AMBER, OpenMM, FEP+BioEmu, AI2BMD ⓚConformational ensembles, binding free energyITC/IC50, stability measurements
MaterialsVASP, Quantum ESPRESSO (DFT)UMA, MACE-MP, MatterSim, Orb, PET-OAM; GNoME, MatterGenCrystal energy / forces → stability, propertiesSynthesis + XRD, property measurement
Chemistry / quantum chemistryORCA, Gaussian, PySCFUMA (trained on OMol25); ether0 (reasoning model)Molecular energies, reactions, propertiesCalorimetry, spectroscopy, yield
Weather / climateECMWF IFS, E3SMGraphCast / GenCast / WeatherNext 3, Aurora, NeuralGCM ⓚEvolution of the atmospheric stateWeather stations, satellites, reanalysis
Fluids / engineeringOpenFOAM, Ansys FluentFNO neural operators, NVIDIA PhysicsNeMo; The Well (15 TB of data)Navier–Stokes flow fieldsWind tunnels, PIV
Fusion / plasmaGyrokinetics GENE/GX, free-boundary equilibrium FGE ⓚTORAX (differentiable transport in JAX) ⓚPlasma equilibrium and transportTokamak diagnostics
Particle physicsGeant4, PythiaCaloChallenge generative models (CaloDiT-2, etc.)Particle–detector interactionsCollider data
Astrophysics / cosmologyIllustrisTNG, SIMBA, ASTRID, EAGLECAMELS suite + neural inferenceLarge-scale structure of the universeSky surveys
NeuroscienceNEURON, Blue Brain (ended 2024) ⓚMouse visual cortex digital twin (MICrONS); FlyWire + NeuroMechFly / flybody; BAAIWorm ⓚNeural responses, brain–body closed loopCalcium imaging, electrophysiology, behaviour
Medicine / physiologyUVA/Padova T1D simulator, cardiac electrophysiology modelsVirtual patient cohorts, digital-twin control armsPatient response to drugs / devicesClinical trials
Robotics / embodied AIMuJoCo, Isaac Sim, Genesis, SAPIENVideo world models (Cosmos, SC3-Eval)Rigid-body contact dynamics, visionReal-robot success rate
Mathematics (control group; not one of the three generations: it is the verifier / ground truth itself)— (Lean 4 and Isabelle are proof checkers, not simulators)AlphaProof, DeepSeek-ProverWhether a proof holdsThe prover kernel itself is the ground truth

Discipline by discipline

2.1 Cell biology: the virtual cell

Representative work

  • Arc Institute State: Arc's first virtual cell model, trained on 100M+ single-cell perturbation profiles (Tahoe-100M, Parse-PBMC, Replogle-Nadig). It predicts the transcriptomic response of stem cells, cancer cells and immune cells to drug, cytokine or genetic perturbations. Arc reports that on Tahoe-100M, State is 50% better than existing models at distinguishing perturbation effects and twice as accurate at identifying differentially expressed genes.
  • Data infrastructure: the Arc Virtual Cell Atlas has passed 600M cells. In Jan 2026 Tahoe, Arc and Biohub launched a new joint dataset project with more than 4× the perturbation richness of Tahoe-100M.
  • A separate route works at the proteome level: An operational perturbation proteomics-based virtual cell model (Nature 2026).

sim2real gap: the largest and most contested of any field

  • Virtual Cell Challenge 2025: more than 5,000 registrants and 300+ teams in the final. First place went to BioMap, the Generalist prize to Altos Labs, and third place to Outlier (TransPert, a statistical model that uses only summary data). The official wrap-up states plainly that prediction models do not yet consistently beat naive baselines across all metrics, with clear progress only on perturbation discrimination and differential-expression detection.
  • Independent benchmark (2026): on unseen single-gene perturbations, the best model (State) beats the "cell mean" baseline by 26%, a tuned linear/additive baseline by 19%, and GEARS by 22%. Several foundation models failed to beat even the linear baseline. Deep models genuinely pull ahead in three places: combinatorial (non-additive) perturbations, transfer across cell types, and representation tasks (annotation, integration).
  • Recommended metrics:
    • Overlap of differentially expressed genes (DE score)
    • Perturbation discrimination: can the prediction be matched back to the perturbation that produced it?
    • Full-spectrum MAE
    • Every metric normalised against the "perturbation mean" baseline
    • The 2025 lesson was that narrow metrics get overfit, so 2026 moves to an aggregate of six metrics, with the cell-eval evaluation tool co-built with NVIDIA.
  • The harder test is prospective validation: the hit rate of model-recommended perturbations in real Perturb-seq or phenotypic screens.
  • VCC 2026 upgrade: the task moves from "generalising within one cell type" to "zero-shot prediction of CRISPRi responses in six entirely new cell lines", with no challenge-specific training set. Deadline Nov 5; results in mid-to-late November.

Relationship to AI

  • The virtual cell is itself an AI model: Generation 3 simulation.
  • It is starting to be used as an environment for agents: BioDiscoveryAgent iteratively proposes perturbations inside an in silico model, simulates outcomes and revises hypotheses. 2026 brought "virtual cell × LLM agent" benchmarks and frameworks such as AssayBench, VCWorld and CellScientist.
  • Risk: if you use a simulator that cannot beat a linear baseline as the reward, the agent learns the simulator's biases.

2.2 Proteins, RNA and genomes

Representative work

  • ESM3 (EvolutionaryScale): a generative language model over three tracks at once: sequence, structure and function. Its flagship example, esmGFP, shares only 58% sequence identity with the nearest known fluorescent protein, which at evolutionary rates corresponds to more than 500 million years of divergence; wet-lab tests confirmed it fluoresces. A third party used it for knotted-protein design with a guided-generation success rate of 89/100 (a computational metric, not wet-lab validated).
  • AlphaFold3 / Boltz: structure prediction covering protein, nucleic acid and ligand complexes.
  • RFdiffusion / BindCraft / AlphaProteo: binder design.
  • Evo 2 (Arc): a genome language model spanning all domains of life.

sim2real gap: here "real" means wet-lab success rate

  • Published binder hit rates: RFdiffusion (2023) about 19%; AlphaProteo (2024) from single digits to 88% depending on target; BindCraft (Nature 2025) 46.3% on average, ranging from 10% to 100%.
  • Discount for production: CROs report that in real projects, unfiltered RFdiffusion/BindCraft output achieves only a fraction of the published wet-lab hit rate, and aggressive developability filtering is needed to recover it. The real lever for raising hit rates from <1% to tens of percent is not bigger models but the wet-lab loop of design → synthesis → measurement → feedback.
  • Genome scale: whole bacteriophage genomes generated with Evo 1/Evo 2 (bioRxiv Sep 2025, later published in Science). About 300 designs were screened, yielding 16 active phages, a success rate of roughly 5%. Some outperformed natural ΦX174 in growth competition and lysis kinetics, and cocktails rapidly broke through resistance in three E. coli strains. This is the first success-rate anchor for whole-genome generation.
  • RNA lags clearly behind proteins: in the CASP16 RNA monomer track, deep learning (AlphaFold3, DeepFoldRNA) only reached the top when combined with templates, physics models and expert manual refinement; pure end-to-end was not enough. The hybrid team GuangzhouRNA-human placed second by Z-score. EnsembleFold achieved a mean TM-score of 0.657 on 19 ensemble targets, 10.2% better than the AlphaFold3 baseline (a TM-score of 0.657 would count as only middling for proteins).
  • Recommended metrics: wet-lab hit rate (how many designs tested per how many made), the distribution of affinity Kd, and developability pass rate; TM-score / RMSD / DockQ for structure; and for pre-screening, the correlation between self-assessed scores such as pLDDT and ipTM and actual success (a 2025 study evaluated the zero-shot predictive power of AlphaFold/ESMFold/ProteinMPNN and concluded they are useful but far from sufficient).

Relationship to AI: the models are AI themselves. The reverse relationship is "structure predictor as the reward for a design model", currently the most mature case of "AI training AI with a simulator", as in ProteinZero and similar online-RL self-improvement work.

2.3 Molecular dynamics and drug binding free energy

Representative work

  • Classical MD: GROMACS, AMBER, OpenMM. Force field plus numerical integration, simulating nanoseconds to milliseconds.
  • FEP+ (Schrödinger) / OpenFE: alchemical free-energy perturbation for computing binding affinity; the current workhorse of computational medicinal chemistry.
  • BioEmu (Microsoft Research, Science 2025): generative deep learning that samples protein equilibrium ensembles directly, producing thousands of statistically independent structures per hour on a single GPU, trained on more than 200 ms of aggregated MD data.

sim2real gap: there is a widely accepted "1 kcal/mol threshold"

  • BioEmu's relative free-energy error against millisecond MD and experimental data is < 1 kcal/mol, with a 4–5 order-of-magnitude speedup; correlations for small-protein folding stability and mutant ΔΔG exceed 0.6. Note this is a mixed evaluation, partly "against MD" and partly "against experiment": a textbook Generation 2 gap.
  • FEP's real gap has an elegant controlled comparison (OpenFE's 2025 large-scale collaborative assessment): on 58 public systems, weighted RMSE 1.73 kcal/mol (10 reached sub-kcal accuracy); on 37 private pharma systems, weighted RMSE 2.44 kcal/mol (only 2 reached sub-kcal). The same methods lose 40% of their accuracy at real-project difficulty. This is the single pair of numbers most worth remembering from this whole report.
  • Recommended metrics: ΔΔG RMSE (kcal/mol), Spearman/Kendall rank correlation (in lead optimisation, ranking matters more than absolute values), and prospective hit-enrichment factor.

2.4 Materials

Representative work

  • Generation 1: DFT (VASP / Quantum ESPRESSO), with the PBE functional as the de facto standard.
  • Generation 2 (MLIPs, universal interatomic potentials): MACE-MP, MatterSim, Orb, SevenNet, Meta's UMA, EPFL's PET-OAM-XL. The proving ground is Matbench Discovery (leaderboards for stability classification, geometry optimisation, phonons, thermal conductivity κ, MD and diatomic potential curves). Models trained on OMat24 already exceed 0.9 stability F1, with formation-energy accuracy of about 20 meV/atom (relative to DFT).
  • Generative: GNoME (DeepMind, 2.2 million predicted stable crystals ⓚ) and MatterGen (Microsoft, Nature 2025).
  • Automated experiments: A-Lab (LBNL) synthesised 41 of 58 targets (about 71%) over 17 days of unattended operation.

sim2real gap: the textbook case of two stacked gaps

  • Layer 2 (MLIP vs. DFT) is already small: F1 > 0.9, 20 meV/atom.
  • Layer 1 (DFT vs. experiment) is the real ceiling: PBE formation energies deviate from higher-level methods on the order of 0.15 eV/atom. Band gaps are the disaster zone: LDA/PBE MAE exceeds 1 eV, with MAPE around 50% (on some benchmarks, PBE has an MAD of 1.28 eV over 70 insulators and PBEsol an MAE of 1.35 eV over 121 materials); HSE06/mBJ bring this down to 0.28–0.30 eV. So even an MLIP that learns DFT to within 1 meV cannot fix DFT's own 1 eV band-gap error. This is the most easily obscured fact in AI for materials.
  • Layer 3 (design → synthesis): MatterGen, conditioned on a 200 GPa bulk modulus, generated TaCr₂O₆, which was successfully synthesised and measured at 169 GPa, a relative error under 20%. However, the structure showed Ta/Cr compositional disorder, and critics argue the compound was already reported in 1972 and falls within training-set coverage. A-Lab's 41/58 has also been questioned by peers over the rigour of its XRD phase identification ⓚ.
  • Note: XRD is not a simulator but the L3 tool for obtaining ground truth. X-ray diffraction is run on the synthesised product to see whether peak positions and intensities match the target structure. Computing the theoretical diffraction pattern from a candidate crystal structure (Bragg's law plus structure factors, Rietveld refinement) is a deterministic physics calculation, but it is an analysis tool, not something AI replaces or trains on. False positives in XRD phase identification (multiphase mixtures, or compositional disorder misread as the target phase) are among the main error sources in the L3 metric "synthesis success rate". "Synthesisability" and "novelty" are two things that existing simulations essentially ignore.
  • Recommended metrics: MAE of energy / forces / stress (against DFT); stability-classification F1 and discovery precision (against the DFT convex hull); property metrics such as κ_SRME; and against experiment, synthesis success rate, relative error on the target property, and whether the compound is genuinely new.

2.5 Chemistry and quantum chemistry

Representative work

  • Generation 1: ORCA, Gaussian, PySCF.
  • OMol25 (Meta, May 2025): more than 100M high-accuracy DFT calculations (ORCA, ωB97M-V/def2-TZVPD) covering biomolecules, metal complexes and electrolytes, consuming 6B+ supercomputer core-hours. It ships with UMA, trained on 30 billion atoms, which uses a Mixture of Linear Experts to ingest data from several different DFT settings at once.
  • Reactions and synthesis: retrosynthesis planning and self-driving labs.
  • ether0 (FutureHouse): a 24B chemistry reasoning model trained with GRPO and verifiable rewards on 640k problems across 375 task types. It outperforms general LLMs, specialised chemistry models and even human experts on molecular design and reasoning, with the biggest edge on open-ended problems.

sim2real gap

  • Quantum chemistry is among the most accurate disciplines for small-molecule energies (the gold standard CCSD(T) reaches chemical accuracy of 1 kcal/mol ⓚ), but accuracy degrades sharply once you move to condensed phases, solvents, reaction kinetics and yields.
  • Synthetic yield is almost impossible to predict from first principles, which is the fundamental reason self-driving labs exist.
  • Recommended metrics: energy MAE (kcal/mol), force MAE, reaction-barrier error; on the experimental side, MAE of predicted yield and the experimental executability rate of retrosynthetic routes.

2.6 Weather and climate

Representative work: ECMWF IFS (Generation 1); GraphCast; GenCast (diffusion-based probabilistic forecasting that beats ECMWF ENS on more than 96% of targets); WeatherNext 3 (Aug 2026, an FGN mesh transformer that assimilates raw satellite observations directly, at 0.05° for stations, 0.1° for the surface and 0.25° for pressure levels, initialised hourly); Microsoft Aurora; and NeuralGCM (a hybrid with a differentiable dynamical core).

sim2real gap: this field supplies the cleanest evidence of benchmark inflation in the whole report

  • AI weather models are almost all trained and evaluated against ERA5 reanalysis, and ERA5 is itself a model product. An observation-based study of the South Asian monsoon found that AI models' MAE against real weather stations is 15–45% higher than their MAE against ERA5 (across all variables, including precipitation). The conclusion: "standard benchmarks centred on reanalysis systematically overestimate operational skill."
  • WeatherNext 3's response, ingesting raw observations directly, is a signal of where the field is heading.
  • Recommended metrics: RMSE / ACC (deterministic); CRPS, ensemble spread-skill ratio and Brier score (probabilistic); extreme events scored separately; and scores must be reported against observations, not just against reanalysis. WeatherBench 2 is the standard venue ⓚ.

2.7 Fluids and engineering CFD

Representative work: OpenFOAM and Ansys Fluent (Generation 1: RANS/LES/DNS); FNO neural operators and NVIDIA PhysicsNeMo; the data infrastructure The Well (PolymathicAI: 15 TB across 16 physical domains, from biological systems, fluids and acoustic scattering to magnetohydrodynamics and supernovae, with a unified PyTorch API and pretrained baselines); and CarBench (2025), a benchmark dedicated to automotive aerodynamics.

sim2real gap

  • Generation 1's own gap comes mainly from turbulence models and boundary conditions. The AIAA Drag Prediction Workshop (DPW) series concluded that at cruise conditions with attached flow, RANS predicts increments between design variants acceptably; once there is large-scale turbulent separation or off-design operation, results from different groups scatter widely, sometimes even producing separation bubbles that do not exist in the wind tunnel. LES/DNS are more accurate but too expensive for routine industrial use.
  • Generation 2 (neural surrogates) is about 1,000× faster than CFD, but its teacher is CFD, so the errors stack: surrogate error plus turbulence-model error.
  • Recommended metrics: relative L2 error of field quantities, energy-spectrum fidelity, long-horizon rollout stability (the main way neural PDE solvers die), and drift in conserved quantities; against physical hardware, deviation of drag/lift coefficients from wind-tunnel data.

2.8 Fusion and plasma

Representative work: free-boundary equilibrium evolution models (FGE, used at TCV), gyrokinetics (GENE/GX), and TORAX (DeepMind's differentiable transport solver in JAX, which supports backpropagation and RL) ⓚ.

sim2real gap: the best example of "train in simulation, deploy straight to hardware"

  • DeepMind × the Swiss Plasma Center (Nature 2022): an RL policy trained only in the simulator was then run zero-shot on the real TCV tokamak, successfully controlling elongated, negative-triangularity and "snowflake" configurations, and even sustaining a "droplet" state with two separate plasmas at once. Follow-up work (2023) improved shape accuracy by 65%, substantially reduced long-term plasma-current error, and cut training time for new tasks by more than 3×.
  • Why zero-shot works here: electromagnetic and equilibrium equations are a strong physical prior, the state is observable, the control frequency is high (kHz), and the cost of failure is bounded. Biology satisfies none of these conditions.
  • Recommended metrics: shape error (cm), current-tracking error, discharge survival rate, and first-attempt success rate on the real machine.

2.9 Particle physics and cosmology

  • Geant4 is the first-principles Monte Carlo of particles passing through detectors, and the entire LHC physics programme is built on it. Data/MC differences are corrected with "scale factors", usually at the level of a few percent ⓚ. Its problem is not accuracy but cost: calorimeter simulation is the most time-consuming part. Hence CaloChallenge 2022: 30+ generative models competed; diffusion and transformer models had high fidelity but were slow, while GANs/VAEs were fast but worse. CaloDiT-2 now agrees well with Geant4 on commonly used physics observables. This is the purest form of "training AI on simulation data to replace the simulation."
  • Cosmology: hydrodynamic simulations such as IllustrisTNG / SIMBA / ASTRID / EAGLE each use different subgrid physics; they are themselves different approximations of reality. The CAMELS project turns this into a methodology: train on TNG, test on SIMBA, and use cross-simulator generalisation to measure sim2real risk. Estimators such as MIEST do representation learning specifically to "erase the simulator's fingerprint." This cross-simulator leave-one-out approach is the evaluation paradigm most worth copying into other disciplines.

2.10 Neuroscience

Representative work

  • Mouse visual cortex digital twin (MICrONS, Nature 2025): a foundation model trained on large-scale calcium imaging from multiple mice watching natural videos. It predicts neuronal responses to entirely new stimulus domains (coherent moving dots, noise), validated in vivo, and can also predict cell types, dendritic features and connectivity. The accompanying electron-microscopy reconstruction covers 200k+ cells and 500 million synapses.
  • Fruit fly: the FlyWire whole-brain connectome (about 139k neurons) plus NeuroMechFly v2 / flybody (87 joints, a 3D body based on X-ray tomography, running on MuJoCo). The Nature 2025 flybody paper achieved whole-body physics-simulated motor control, and 2026 brought work such as FlyGM, in which a connectome graph model drives whole-body locomotion.
  • The nematode has BAAIWorm ⓚ; the Blue Brain Project ended in 2024 ⓚ.

sim2real gap

  • The sensory-encoding layer is already quite accurate (digital-twin predictions for new stimuli have been validated in vivo), but behaviour, learning and cross-region closed loops remain far off.
  • The field has recently suffered from "viral demo" pollution (videos like "a fruit fly playing Beat Saber"); check whether a demo really is a closed loop constrained by the connectome.
  • Recommended metrics: single-trial response correlation / explained variance, generalisation to out-of-distribution stimuli (the key one), and agreement of predicted connectivity with EM ground truth; on the behavioural side, statistical agreement of kinematics with real fly gaits.

2.11 Medicine and physiology: the most underrated success story

  • The UVA/Padova type 1 diabetes simulator: accepted by the FDA in 2008 as a substitute for animal trials in the preclinical design and safety testing of artificial-pancreas controllers. Since then, animal experiments for designing artificial-pancreas algorithms have all but disappeared. The simulator contains 300 virtual subjects (100 adults, 100 adolescents, 100 children). By 2023, six algorithms developed on it had entered clinical use: Medtronic 670G/770G, 780G, CamAPS FX, Diabeloop, Tandem Control-IQ and Insulet Omnipod 5. This is the strongest precedent for "a simulation environment directly yielding regulator-accepted real products", and it is a hand-built ODE model, not deep learning.
  • Regulatory framework: the FDA has issued final guidance on assessing the credibility of computational modelling, based on ASME V&V 40, which defines eight categories of credibility evidence (bench validation, in vivo validation, population validation and so on). Its known limitations: V&V 40 assumes a single validation comparator and its language leans towards bench testing, so it does not fully fit in silico clinical trials.
  • Cardiology: virtual patient cohorts plus digital twins that infer activation and repolarisation sequences from ECGs, used in device development.
  • Recommended metrics: distributional overlap between virtual and real cohorts on endpoints, the V&V 40 credibility level for a given context of use (CoU), and the hardest one of all: whether regulators accept it.

2.12 Robotics and embodied AI (tracked separately, so brief here)

  • Generation 1: MuJoCo (the gold standard for contact), Isaac Sim, Genesis, SAPIEN. Generation 3: video world models (Cosmos, SC3-Eval, WorldEval).
  • The sim2real gap is measured differently here, mainly as ranking fidelity. SIMPLER used 1,500+ paired sim-vs-real evaluations to show strong correlation is achievable. The key metrics are Pearson r and MMRV (mean maximum rank violation), because when several policies have similar real-world scores, r becomes misleading; the real question is "did the simulator get the ranking wrong?"
  • Findings so far: pure physics-simulation benchmarks correlate only weakly with real robots (r ≈ 0.4), while video world-model evaluation can reach 0.9+. Absolute real-robot success rates are typically only ~44%, in sharp contrast to inflated simulation numbers.

2.13 Mathematics: the control group where verification-layer gap ≈ 0 (but kernels have bugs)

Strictly speaking, Lean 4 / Isabelle are not simulations but ground truth itself: if the kernel accepts it, the proof holds, and there is no approximation in the verification step. That is exactly why the AlphaProof, DeepSeek-Prover-V2 and Leanabell-Prover-V2 line has progressed fastest: verifiable reward is unlimited, noiseless and free. The new direction in 2026 is plugging the Lean REPL into RL training for dense, tactic-level credit assignment, rather than rewarding only the final outcome.

Lean's metatheory (why it counts as ground truth)

  • The foundation is dependent type theory, specifically a variant of the Calculus of Inductive Constructions (CIC). Under the Curry–Howard correspondence, propositions are types and proofs are terms, so "checking a proof" is the same as "type checking".
  • Universe hierarchy: Prop (impredicative and proof-irrelevant: any two proofs of the same proposition are definitionally equal) ⊂ Type 0 ⊂ Type 1 ⊂ …, with universe polymorphism.
  • Inductive types (including mutual and nested ones) get recursors generated automatically by the kernel. Structural and well-founded recursion are compiled to recursors by the front end, so the kernel contains no termination checker. Coq's guard checker has had several historical soundness holes; Lean avoids this class of problem by design.
  • Beyond the type theory, only three axioms are added: propext (propositional extensionality), Quot.sound (quotient types, from which function extensionality follows) and Classical.choice (the axiom of choice, from which excluded middle follows via Diaconescu's theorem). So Mathlib is classical mathematics.
  • Consistency: Carneiro (2019, The Type Theory of Lean) built a set-theoretic model showing Lean is consistent relative to ZFC plus countably many inaccessible cardinals ⓚ. In other words, trusting Lean ≈ trusting a set theory slightly stronger than ZFC. By Gödel's second incompleteness theorem, this is the best that can be done.
  • Known theoretical blemishes: definitional equality is not transitive, type checking is undecidable in theory, and subject reduction fails ⓚ. These affect completeness (the kernel may reject proofs it should accept), not soundness (whatever it accepts is correct).

Where things have actually gone wrong in the trust chain (ordered from most to least trustworthy)

LayerCan it be wrong?History
MetatheoryCannot prove its own consistency; consistent relative to ZFC + inaccessible cardinalsNo flaw ever found
Kernel implementation (a few thousand lines of C++)Yes, ordinary software bugsJul 2026, #14576: an AI-assisted Collatz "disproof" exploited a missed check on phantom parameters of nested inductive types to prove False; fixed within an hour. The independent kernels nanoda (which had a different bug) and Lean4Lean (same bug) both failed to catch it. OpenAI's Selsam then used AI auditing to dig up several more, all of which nanoda could catch.
native_decide / reduceBoolYes: trusts an unverified compilerIn 2023 Carneiro constructed a contradiction using random IO; the fix made it an explicit axiom, trustCompiler, visible via #print axioms.
Front-end elaborator / tacticsBugs are common, but the kernel re-checks, so they are mostly harmless—
Formalisation gapThe most common: the Lean statement ≠ the mathematical statement you meant to proveMis-formalised problems have repeatedly been found in benchmark sets

Other provers are no different: Rocq (Coq) maintains a dedicated critical-bugs.md recording several historical holes that could prove False (Set+1 type computation, native_compute constant-name collisions, the guard checker and more). The more expressive the system, the larger the kernel and the larger the attack surface: HOL Light's kernel is only a few hundred lines ⓚ, while Lean's and Rocq's are much larger.

Implication for RL environments: RL is extremely good at finding back doors, so kernel soundness is now an AI-safety infrastructure problem. Engineering discipline:

  • Ban native_decide, custom axiom declarations and direct metaprogramming declarations; whitelist with #print axioms.
  • Once a proof passes, re-check it with at least one independent kernel (such as nanoda or lean4checker).
  • When something "proves" a famous conjecture or an anomalous result, treat it as a bug by default.

The long-term direction is verified verifiers: Lean4Lean formalises Lean's type checker in Lean; Rocq has MetaCoq, and HOL Light has Candle, built on CakeML ⓚ.

This control group explains the structural regularity of the whole report: how fast AI advances in a field is proportional to how verifiable its reward is, and inversely proportional to its sim2real gap.

Comparing sim2real gaps: a metrics framework

3.1 The three layers must be reported separately

LayerQuestion it asksTypical metricsCommon cheats / self-deceptions
L1 simulator vs. realityIs the physical fidelity good enough?MAE / relative error against experimental measurementsNot measuring it at all and assuming the simulator is right
L2 AI emulator vs. simulatorHas it learned the teacher well?Energy/force MAE, L2 field error, rollout stabilityPassing off pretty L2 numbers as L1
L3 design → realityDoes what it produces actually work?Wet-lab hit rate, synthesis success, real-robot success, clinical endpointsReporting only the best targets, omitting the denominator

3.2 Scoring the current gap in each field

FieldL1 (simulation vs. experiment)L3 (design → reality)Best metric to use
Mathematics / formalisationNone (it is the ground truth; residual risk lies in the kernel implementation and the formalisation gap)Verifiable, but needs multi-kernel re-checkingIndependent-kernel re-check pass rate + axiom whitelist
Plasma controlSmall; enough for zero-shot on hardwareSucceeded at TCVShape error, first-attempt success on hardware
Robot manipulationMedium (vision + contact)Real robot ~44%MMRV + Pearson r
WeatherSmall, but 15–45% worse against observationsOperational forecasts already liveCRPS, RMSE against observations
CFDMedium (wide scatter for separated flow and off-design)Usable in engineeringRelative L2, drag-coefficient deviation
Particle physicsSmall (data/MC within a few percent)—Agreement of physics-observable distributions
MaterialsLarge (band-gap MAE >1 eV)Synthesised + property error <20% (single case)Synthesis success rate, relative property error
Quantum chemistry / small moleculesSmall (~1 kcal/mol)Yield cannot be predictedEnergy MAE, yield MAE
Drug binding free energyMedium → large (1.73 → 2.44 kcal/mol)Prospective enrichment factorΔΔG RMSE + Spearman
Protein designNot applicable5% (whole genome) to 46% (binders)Wet-lab hit rate (with denominator)
NeuroscienceSmall for sensory encoding, large for behaviour—Explained variance on OOD stimuli
Medicine / physiologyMedium, but accepted by regulatorsSix marketed products alreadyV&V 40 credibility level
Virtual cellsLargestNo consensus yet on prospective validationMultiple metrics normalised to the mean baseline

3.3 General evaluation discipline (cross-cutting lessons from this survey)

  1. Always beat the naive baseline first. In virtual cells, the "take the mean / linear additive" baselines have already eliminated a batch of foundation models. For any scientific model claiming SOTA, first ask: "how much better than the average is it?"
  2. Evaluate against real observations, not against simulation products. The 15–45% gap in weather is the price of not doing so.
  3. Report public benchmarks and private / real tasks separately. FEP's 1.73 vs. 2.44 is the cleanest example.
  4. Cross-simulator leave-one-out validation. The CAMELS approach (train on TNG, test on SIMBA) should be standard in every field that trains on simulated data.
  5. Any single metric will be overfit. VCC 2025's lesson is written into the official wrap-up: "A narrow scoring surface induces optimisation for the metric rather than for the biology."

Five ways simulation relates to AI, in decreasing maturity

Form 1: simulation generates data to train AI (most mature, already industrialised)

  • OMat24 / OMol25 → MLIPs: OMol25 contains 100M+ DFT calculations and burned 6B+ supercomputer core-hours; UMA was trained on 30 billion atoms. This is "using supercomputers to distil the laws of physics into a neural network."
  • ERA5/HRES → GraphCast/WeatherNext; 200 ms of MD → BioEmu; Geant4 → CaloDiT; The Well's 15 TB → neural PDE solvers.
  • The common pattern: Generation 1 simulators have become data factories, and AI captures a 10³–10⁵× speedup.

Form 2: AI replaces simulation (in progress)

The key criterion is whether the replacement's error exceeds the original simulator's error against reality. Weather has crossed that line (AI is more accurate than IFS). Materials MLIPs have crossed the line on the need for DFT, but not DFT's own ceiling. CFD has not crossed it yet.

Form 3: simulation as an RL environment for control policies (real successes already)

  • TCV tokamak: trained in simulation → zero-shot on the real machine.
  • UVA/Padova T1D: trained in simulation → FDA acceptance → six marketed products.
  • Robotics: massive domain randomisation in Isaac/MuJoCo, but ranking fidelity remains the bottleneck.
  • What they share: observable state, strong physical priors, high action frequency, and failures that can be rolled back. All three successes are control problems, not discovery problems.

Form 4: simulators and predictors as tools and rewards for LLM agents (the main battlefield of 2025–2026)

  • ether0: deterministic chemistry checkers such as RDKit serve as the verifier; trained with GRPO on 640k problems across 375 tasks.
  • Lean: a near-noiseless verifier with dense tactic-level rewards. But the Jul 2026 #14576 incident proved that RL/AI will actively exploit kernel bugs, so multi-kernel re-checking is needed.
  • Virtual-cell agents: BioDiscoveryAgent, CellScientist, VCWorld, AssayBench.
  • Structure predictors as the reward for binder design (online RL such as ProteinZero).
  • The chronic disease of this form is reward hacking: the weaker the verifier, the more the agent games it. This has been quantified for coding RL environments: in a sample of SWE-bench Verified, 28.5% of tasks had tests weak enough that a wrong patch still passed. Scientific verifiers will only be weaker.

Form 5: nature as the RL environment (the most expensive, and the capital consensus of 2026)

  • Periodic Labs: builds its own physical laboratories. Models first try ideas in simulators of progressively higher fidelity, then move to real experiments, with reward coming from the experimental outcome itself. In their framing: "ChatGPT optimises against a reward model of human preferences; Periodic optimises against experiments."
  • Lila Sciences: an autonomous-lab platform for life sciences, chemistry and materials, running a hypothesis → experiment → learning loop.
  • OpenAI is already running closed-loop experiments with pharma companies (GPT-5 iterating protocols autonomously); DeepMind opens an automated materials lab in 2026; Anthropic is taking the enterprise-tools route (Claude for Life Sciences, connecting to Benchling, 10x and PubMed).
  • The hard economic constraint (as summarised by SemiAnalysis): a single biology experiment costs hundreds to thousands of dollars and takes hours, whereas grading a coding task is nearly free. Data return per dollar is several orders of magnitude lower, and on top of that biology rollouts are long, rewards are sparse and the literature contradicts itself. The industry's mitigations are step-level rubric rewards (rewarding the process, not just the outcome) and in-flight weight updates (roughly 2× the iterations in the same time). For now, rewards in scientific domains are still mostly human grading.

The verdict in one sentence

A simulation environment's value to AI depends on how close it is to a verifiable reward. Lean is closest (verification-layer gap ≈0, but you must guard against kernel back doors); control simulations come next (small gap, zero-shot possible); materials and molecules sit in the middle (gap known and correctable); virtual cells are furthest (gap unknown, baselines not yet beaten). That is why the best-funded players have chosen to bypass simulation and buy real experiments as their reward.

Implications for AI4S platforms and founders

  1. If you are building "scientific RL environment" products, prioritise by verifier strength, not by how hot the discipline is. Under-served areas with strong verifiers: formal mathematics, deterministic checks in quantum chemistry, crystal-structure validity, executability of retrosynthetic routes, and reproducibility of bioinformatics pipelines. Virtual cells are too early to serve as a reward.
  2. Selling "cross-simulator / cross-baseline evaluation" is itself a product. CAMELS-style leave-one-out validation, VCC's mean-baseline normalisation, SIMPLER's MMRV: this evaluation discipline is missing in most disciplines, and wherever it is missing there is a trust gap.
  3. Passing off L2 as L1 is the most common marketing line in this industry. Whenever a partner claims "we're more accurate than DFT/MD", ask whether the comparator is a simulator or an experiment.
  4. The T1D simulator is the commercial precedent most worth studying: a hand-built ODE model that, because it won regulatory acceptance, has dominated a market for nearly 20 years and spawned six marketed products. The moat of regulatory acceptance runs far deeper than model accuracy.

Appendix A: the industry behind Generation 1 simulation software

Open source vs. commercial, the companies, their market caps and profits.

Conclusions

  1. Two worlds. Engineering simulation (CFD / finite elements / multiphysics) is highly commercialised and oligopolistic, with operating margins of 30–46%. Scientific simulation (DFT / MD / NWP / Geant4 / MuJoCo) is dominated by open-source and academic software, subsidised by governments, international organisations or big tech. Schrödinger, the only listed pure scientific-computing company, has a market cap of about $2B and has long been loss-making.
  2. In 2024–2026 engineering simulation was bought up by the EDA giants and Siemens, about $50B in total: Synopsys acquired Ansys (about $35B), Cadence acquired BETA CAE and Hexagon's MSC (about €2.7B), and Siemens acquired Altair (about $10B).

A.1 By field: open source or commercial, and who backs it

FieldMainstream softwareOpen / commercialInstitutions / companies behind it
CFD / FEA / multiphysics (engineering)Ansys Fluent/Mechanical, Simcenter STAR-CCM+, Altair, Abaqus, MSC Nastran, COMSOL, MATLAB/SimulinkMostly commercialSynopsys (Ansys), Siemens (Simcenter + Altair), Dassault (SIMULIA), Cadence (BETA CAE, MSC, Fidelity), COMSOL and MathWorks (private)
Open-source engineering alternativesOpenFOAM, SU2, code_saturneOpen sourceOpenFOAM Foundation / OpenCFD (ESI, acquired by Keysight in 2023 ⓚ)
DFT / quantum chemistryVASP, Gaussian, Materials Studio/CASTEP, QuantumATK, ORCA, Q-ChemCommercial or academic licenceVASP charges per research group; Gaussian Inc. (private; $12k annual maintenance, long controversial for barring developers of competing software); Dassault BIOVIA; Synopsys; ORCA free for academics, paid for commercial use ⓚ
Open-source quantum chemistryQuantum ESPRESSO, CP2K, ABINIT, PySCF, Psi4, NWChemOpen sourceAcademia and national research funding
Molecular dynamics / drugsGROMACS, OpenMM, LAMMPS, NAMDOpen source (the real workhorses)Academia, NIH, DOE national labs
AMBER, CHARMMSemi-open (AmberTools is open source; the GPU engine is paid ⓚ)Academic consortia
Desmond, FEP+, OpenEyeCommercialDeveloped by D.E. Shaw Research, distributed by Schrödinger; OpenEye acquired by Cadence in 2022 ⓚ; OpenFE is open-source FEP funded by a pharma consortium
Numerical weather / climateECMWF IFS, NOAA UFS/GFS, WRF, CESM, E3SMPublic / open sourceGovernments and international organisations; OpenIFS open-sourced under Apache 2.0 on Mar 5, 2026
Particle physicsGeant4, PythiaOpen sourceInternational collaborations led by CERN
RoboticsMuJoCo, Isaac Sim/Isaac Lab, Newton, GenesisAll open sourceDeepMind (acquired MuJoCo in 2021, open-sourced it in 2022 ⓚ); NVIDIA (Isaac Sim under Apache 2.0, Isaac Lab under BSD-3); Newton co-built by DeepMind, Disney and NVIDIA, hosted by the Linux Foundation

Robotics simulators are open source not because they lack value, but because the value is captured by NVIDIA (selling GPUs) and Google (ecosystem).

A.2 Company market caps and profits (Sep 2026)

CompanySimulation assetsMarket capRelevant revenueMargin
SynopsysAnsys (about $35B, closed Jul 17, 2025)Sources disagree widely ($60–96B); share price about $385 on Sep 18FY26 guidance $9.69–9.74B, of which Ansys about $2.98BQ3 FY26 non-GAAP operating margin 41.6%, GAAP 14.4% (heavy amortisation)
Ansys (pre-acquisition)Fluent, Mechanical, HFSS—$2.545B in 2024GAAP 28.2%, non-GAAP 45.7%
CadenceBETA CAE (2024), Hexagon MSC (Feb 23, 2026), Fidelity, OpenEyeAbout $77–78B$5.30B in 2025; 2026 guidance $5.9–6.0B (simulation not broken out)Non-GAAP about 45%
Dassault SystèmesSIMULIA, BIOVIAAbout $34B€6.24B in 2025 (simulation not broken out)Non-IFRS 32.0%; 2026 growth guidance only 3–5%
SiemensSimcenter + Altair (about $10B, Mar 2025)Group market cap not comparableAltair $665.8M in 2024In Q4 FY25, Altair and Dotmatics integration dragged profit by €135M
SchrödingerFEP+, Maestro, DesmondAbout $2.17B$256M in 2025 (software $200M)2025 net loss $103M; Q2 2026 book profit of $6.0M came mainly from a $45.9M equity-investment gain, and the core business is still loss-making
MathWorks (private)MATLAB/Simulink—About $1.5B in 2024Not disclosed
COMSOL (private)COMSOL Multiphysics—On the order of $100M (third-party estimate, low confidence)Not disclosed

Market size: CAE was about $12–13B in 2025, with the top five holding 50–55%; quantum-chemistry software was about $0.62B in 2024, roughly a 20× difference.

A.3 Four judgements

  1. The money is in engineering, not science. Product-certification processes in aerospace, automotive and chips are tied to specific software, so switching costs are very high. Users of scientific simulation want public, reproducible code and naturally lean towards open source.
  2. EDA is swallowing simulation to go "from chip to system". Advanced packaging, thermal management, electromagnetics and signal integrity for AI chips all need multiphysics simulation. The roughly $35B price for Ansys shows mature engineering simulation is priced at more than 10× revenue.
  3. Schrödinger is the cautionary tale. Selling scientific-computing software to pharma hits a ceiling at about $200M in annual software revenue, leaving the company to tell its story through an in-house drug pipeline while staying loss-making. Selling scientific simulation software by itself rarely becomes a big business.
  4. What this means for Generation 2 AI emulators. In engineering, AI surrogates threaten high-margin incumbents, but those incumbents are already integrating AI themselves (Ansys SimAI, PhysicsNeMo partnerships) and hold the certification barrier. In science, AI models (UMA, MACE, GraphCast) are mostly open source too; the software itself is worth little, and the value lies in data, compute and wet-lab validation.

Appendix B: the market for Generations 2 and 3

Companies, funding, valuations and revenue.

Conclusions

  1. Private valuations are far above public-market pricing. Top private valuations: Isomorphic (reported at $15–20B), Lila ($8.5B pre-money, in negotiation), Periodic Labs ($7.5B, in negotiation), CuspAI ($2.6B), PhysicsX ($2.4B). Listed AI-drug-discovery / AI4S companies (Recursion about $1.8B, Schrödinger about $2.2B, XtalPi around HK$30B) are stuck at a few billion dollars, and most are loss-making.
  2. Almost no company makes money by "selling the model". Generation 2 companies sell to engineering and energy customers (the same buyers as Ansys). Generation 3 companies either develop their own pipelines (Isomorphic, Xaira, Recursion), run data factories (Tahoe), or are non-profits that open-source their models (Arc, Biohub). EvolutionaryScale (ESM3) was acquired by the non-profit Biohub in Nov 2025, a sign that the business of independently selling protein models did not work out.
  3. The biggest money is flowing to "real laboratories as the RL environment" (Lila, Periodic Labs), consistent with Form 5 in the main text.
  4. Big tech acts as a platform rather than a software vendor: NVIDIA (Earth-2, PhysicsNeMo, BioNeMo, Cosmos, ALCHEMI) monetises through GPUs, and NVentures has also invested in Lila, Luminary and Orbital. Google DeepMind (WeatherNext, AlphaFold/Isomorphic), Microsoft (MatterGen, MatterSim, Aurora, BioEmu) and Meta (UMA, open source) do not disclose revenue from this work separately.

B.0 Generation 2 vs. Generation 3: differences, examples and capital profile

Generation 2: AI emulators / surrogatesGeneration 3: data-driven world models
TeacherPhysics simulators (DFT, MD, CFD, NWP)Experimental data (sequencing, structure databases, imaging)
Data sourceCompute traded for data: unlimited and noiselessWet labs: expensive, scarce, with batch effects
How to checkRe-run the original simulatorOnly by running another experiment
Accuracy ceilingEqual to the teacherNo clear ceiling, and no floor guarantee either
Selling point10³–10⁵× faster; essentially cost savingDoing what could not be computed before; essentially a new capability
As an RL rewardUsable and correctableCurrently too noisy
Representative modelsMACE, UMA, MatterSim, Orb; GraphCast, GenCast, WeatherNext, Aurora; BioEmu; FNO/PhysicsNeMoArc State, Xaira X-Cell; ESM3, AlphaFold3; Evo 2; mouse visual cortex digital twin
Representative valuationsClustered at $1–3B: CuspAI $2.6B, PhysicsX $2.4B, Tomorrow.io >$1B, DP Technology unicornPolarised: Isomorphic reported at $15–20B, Xaira about $1.3B raised in total, Chai $1.3B; listed Recursion about $1.8B, XtalPi around HK$30B
Revenue shapeSoftware/service revenue from engineering, energy and materials customers: recurring but small, mostly undisclosedMainly a few large pharma partnership deals, highly volatile (XtalPi +201% in 2025, then down in H1 2026)
Valuation anchorThe CAE market (PhysicsX $2.4B ↔ Ansys annual revenue of about $3B)Option value of a drug pipeline; can have almost no revenue
Main riskIncumbents (Ansys, Siemens) ship their own AI and hold the certification barrierThe pricing gap between $10B+ private and $2–4B public valuations

Hybrids: NeuralGCM (physics core plus learned modules), BioEmu (MD plus experimental data), WeatherNext 3 (starting to ingest observations directly). Lila ($8.5B pre-money) and Periodic Labs ($7.5B), which straddle both generations by using Generation 2 models to pre-screen in simulators and real experiments as the training signal, are the category whose valuations are rising fastest.

B.1 Generation 2: AI emulator / surrogate companies

AreaCompanyLatest round / valuationRevenue / notes
Engineering CFD / structuresPhysicsX (London)Series C $300M in Jun 2026 (led by Temasek) at a $2.4B valuation; $489M raised in totalNot disclosed; valued at about $1B in Nov 2025, a 2.4× jump in seven months
Neural Concept (Switzerland, EPFL spin-off)Series C $100M in Dec 2025 (led by Goldman Sachs)50+ customers (GM, GE Vernova, Safran)
Luminary Cloud (US)Series B $72M in Sep 2025; $187M raised in totalNVentures participated
Incumbents with built-in AIAnsys SimAI, Siemens Simcenter AIThe main threat to startups
Materials / chemistry MLIPsCuspAI (Cambridge)Series B $450M in 2026 at a $2.6B valuation (led by KP and NEA)Not disclosed; launched an "AI Materials Foundry"
Orbital Industries (formerly Orbital Materials)Series B $50M in May 2026Open-sourced the Orb potential
DP Technology (深势科技, Beijing)Series C of more than RMB 800M in Dec 2025; unicornDeePMD, Uni-Mol, Hermite, the Bohrium research platform; not listed
WeatherTomorrow.ioSeries F $175M + $35M in 2026; valuation >$1BBuilds its own weather satellites; AI forecasting
WindBorne SystemsSeries B $37M at a $250M valuationRevenue tripled over the past year (its own radiosonde-balloon network plus AI models)
Jua / Brightband$21.5M / $10MEnergy trading / Earth-system AI
Robotics simulationGenesis AISeed round $105M in Jul 2025In-house physics engine plus robotics foundation model
Big techGoogle (WeatherNext), NVIDIA (Earth-2, PhysicsNeMo), Microsoft (Aurora, MatterSim, BioEmu), Meta (UMA)—Monetise via GPUs / cloud, or open-source for free

B.2 Generation 3: data-driven world-model companies

AreaCompanyLatest round / valuationRevenue / P&L
Structure / drug foundation modelsIsomorphic Labs (Alphabet)Series B $2.1B in May 2026 (led by Thrive); valuation reported at $15–20BPartnerships with Novartis, Lilly and J&J; revenue not disclosed
XairaAbout $1.3B raised in totalReleased the X-Cell virtual cell model (4.9B parameters)
Chai Discovery (OpenAI-backed)Series B $130M in Dec 2025 at a $1.3B valuation; >$225M raised in total; reportedly negotiating a new $400M roundNot disclosed
EvolutionaryScale (ESM3)Seed $142M → acquired by Biohub in Nov 2025; about 50 people joinedIndependent commercialisation ended
Virtual cellsTahoe Therapeutics$30M at a $120M valuationData factory (Tahoe-100M); co-building datasets with Arc/Biohub
NoetikNot detailedMultimodal patient-tissue data
Arc Institute, BiohubNon-profitState, Evo 2 and ESM3 open-sourced
Listed companiesRecursion (RXRX)Market cap about $1.8B (Aug 2026)2024 revenue $58.8M; net loss of $536.6M in the first three quarters of 2025
XtalPi (晶泰控股, 2228.HK)Market-cap sources disagree (HK$3.2B vs. HK$33.2B; the latter matches a share price around HK$6.93)2025 revenue RMB 803M (+201%), first profit of RMB 135M; but H1 2026 revenue about RMB 390M (−24%) with a loss of about RMB 230M (secondary summary, to be verified)

B.3 Straddling both generations: the real laboratory as the environment

CompanyFunding / valuationNotes
Lila SciencesSeries A totalling $350M in Oct 2025 at a valuation >$1.3B; reported in Jun 2026 to be negotiating $2B at $8.5B pre-money (led by CalPERS and NVentures)Autonomous labs covering life sciences, chemistry and materials
Periodic LabsSeed $300M in Sep 2025 at a $1.3B valuation; reported in May 2026 to be negotiating $500M at a $7.5B valuationAlready has paying customers in semiconductors; models first try ideas in simulators of rising fidelity, then move to real experiments

B.4 The institutions behind the representative Generation 2 models

ModelFirst releaseInstitutionFoundedParent market cap / valuationParent annual revenueCommercialisation and revenue
MACE (incl. MACE-MP-0)2022 (universal potential 2023)Csányi group, University of Cambridge + ACEsuit open-source communityAcademic——Code and the MACE-MP-0 series are MIT; newer models such as MACE-OFF, OMAT, OMOL, MATPES and MH use an Academic Software License (ASL) that bans commercial use. No direct revenue
UMAMay 2025Meta FAIR Chemistry (grew out of the Open Catalyst Project with CMU, 2020)FAIR 2013; Meta 2004Meta about $1.70T (Sep 18, 2026)$201.0B in 2025, net income $60.5BDownload on Hugging Face by request after agreeing to the FAIR Chemistry License (commercial use allowed, with geographic and use restrictions); not charged for separately
MatterSimMay 2024Microsoft Research AI for ScienceJul 2022 (founded by Chris Bishop); Microsoft 1975Microsoft about $3.71TFY26 (to Jun 2026) $331.8B, net income $133.7BMIT open source, integrated with Azure / Microsoft Discovery; revenue not broken out
Orb2024Orbital Industries (formerly Orbital Materials, London)2022 (founder Jonathan Godwin, ex-DeepMind)Valuation undisclosed; about $71M raised (Series A $16M Feb 2024, Series B $50M May 2026)Not disclosedApache 2.0 open source; the company has pivoted to monetising industrial hardware such as data-centre cooling (Orbital IT)
GraphCast / GenCast / WeatherNext2023 / 2024 / WeatherNext 2 in Nov 2025, WeatherNext 3 in Aug 2026Google DeepMindDeepMind 2010, acquired by Google 2014; Alphabet 2015Alphabet about $4.10T$402.8B in 2025, net income $132.2BData on BigQuery / Earth Engine, early access on Vertex AI; integrated into Search, Gemini, Pixel and the Maps Weather API (indirect monetisation, not broken out)
AuroraJun 2024 (Aurora 1.5 later)Microsoft Research AI for ScienceAs aboveMicrosoft about $3.71TAs aboveHosted deployment on Microsoft Foundry, Planetary Computer Pro; used in MSN Weather; not broken out. Core authors left in Jun 2024 to found Silurian (about $6M raised)
BioEmuPreprint Dec 2024, Science 2025Microsoft Research AI for ScienceAs aboveAs aboveAs aboveMIT open source; no direct revenue
FNO (Fourier Neural Operator)2020Caltech (Zongyi Li, Anima Anandkumar)Academic——The neuraloperator library is MIT open source (2020), absorbed into NVIDIA's and other frameworks
PhysicsNeMo (formerly Modulus)Modulus open-sourced 2023, renamed 2025NVIDIA1993NVIDIA about $5.58TFY26 (to Jan 2026) $215.9B, net income $120.1B; data centre $197.3BApache 2.0 open source; monetised through GPUs and NVIDIA AI Enterprise

Key points

  1. Of these 10 representative models, 8 come from big-tech labs or universities; only Orb is backed by an independent startup (plus Silurian, founded by the departing Aurora team).
  2. The models themselves earn almost no direct revenue: most are open source, and big tech monetises them by driving cloud usage (Azure, Google Cloud), GPU sales (NVIDIA) or improving its consumer products (Google Search and Maps weather, MSN Weather). Against parent revenues of $200–400B a year, this slice cannot be, and is not, disclosed separately.
  3. The companies actually turning Generation 2 models into a business are downstream: PhysicsX, CuspAI, Neural Concept, DP Technology and others, which often build industry-specific products directly on these open-source models and earn money from industry delivery and data.
  4. This means the Generation 2 "model layer" has been commoditised by big-tech open-sourcing, and moats can only be built on industry data, workflow integration and customer delivery.

B.5 Judgements

  1. Generation 2's valuation anchor is the CAE market: PhysicsX's $2.4B maps to Ansys's roughly $3B annual revenue, a bet that AI surrogates will eat part of the traditional simulation budget. The risk is that incumbents ship their own AI and hold the certification barrier.
  2. Generation 3's valuation anchor is the drug pipeline, not the model: Isomorphic's and Xaira's valuations come from options on future drugs. Public markets give "AI platform + pipeline" companies such as Recursion and XtalPi only $2–4B, and their revenue depends on a handful of large partnership deals and swings sharply (XtalPi boomed in 2025 and fell back in H1 2026).
  3. Private markets are putting the highest price on "data + real validation": Lila's and Periodic's valuations rose 5–6× within half a year, while a pure model company (EvolutionaryScale) was absorbed. This matches the main conclusion: the larger a field's simulation gap, the more value migrates towards data and the wet lab.

Appendix C: adoption, open source, market size and closing the Gen-1 gap

C.1 Why universities and companies use Generation 2 instead of Generation 1

The core reason: provided accuracy is "good enough", Generation 2 cuts compute cost by 3–5 orders of magnitude, making a lot of previously impossible research feasible. In practice, though, it does not replace Generation 1; the two are chained together.

  1. 10³–10⁵× faster, which changes the scale of research outright.
    • Materials / chemistry: DFT takes minutes to hours per structure, while universal interatomic potentials (MACE, UMA and others) take milliseconds with near-DFT accuracy. DFT molecular dynamics used to be limited to a few hundred atoms for tens of picoseconds; now tens of thousands to millions of atoms can be run for nanoseconds. Systems that must be large and long-lived, such as interfaces, defects, electrolytes and catalyst surfaces, finally become computable.
    • Weather: GraphCast produces a 10-day forecast in under a minute on a single TPU v4; ECMWF's HRES takes about an hour on a supercomputer.
    • Proteins: BioEmu generates thousands of independent conformations per hour on one GPU. Conformational changes such as cryptic pockets, which used to require millisecond-scale MD to observe, can now be obtained in hours.
  2. Large-scale screening becomes possible. Generation 1 can only compute the few candidates you have already thought of; Generation 2 can sweep millions first and hand a few to DFT or experiment. GNoME predicting 2.2 million crystal structures, 380,000 of them predicted stable, is exactly this pattern.
  3. Probabilistic forecasts and uncertainty get cheap. Ensemble forecasting used to mean running the whole physics model dozens of times; models like GenCast sample dozens to hundreds of members in one go.
  4. Differentiable, so they plug into optimisation and AI pipelines. They support inverse design, can serve as rewards or constraints in RL, and can be called by LLM agents. Legacy Generation 1 Fortran/C code is hard to make differentiable.
  5. Low barriers to entry. Hardware: one GPU is enough, no queueing for supercomputer allocations. Licences: VASP charges per research group, Gaussian's annual maintenance is $12k and commercial CAE charges annual per-seat fees, whereas MACE-MP-0, MatterSim, Orb, BioEmu and PhysicsNeMo are all free and open source. People: Generation 1 software needs experts in functional choice, convergence parameters and meshing; Generation 2 is basically pip install plus a few lines of Python.
  6. Companies value iteration speed. The design loop goes from "change a version and wait overnight for CFD" to "results in seconds", so hundreds of variants can be tried in a day.

Why Generation 1 has not been retired: the standard practice is a three-stage funnel.

Generation 2 coarse screening at scale (millions of candidates; fast and cheap) → keep the top 1% → Generation 1 precise validation (DFT / MD / CFD; expensive but trustworthy) → keep the few most promising → experiment / wind tunnel / clinic as the final judge.

  • Generation 2's training data is computed by Generation 1, so Generation 1 is its accuracy ceiling; for new chemical systems you still have to go back and run DFT/CFD.
  • Out-of-distribution results are untrustworthy: for unseen element combinations, extreme conditions or unseen flow regimes, Generation 2 may give plausible-looking wrong answers.
  • Certification: certification processes in aerospace, automotive, nuclear power and medical devices recognise Generation 1 commercial software plus validation records.
UniversitiesCompanies
Primary motivationStudy scientific questions at larger scales and longer timesShorten design cycles
Typical useFine-tune open-source interatomic potentials on their own systems; study mechanismsAccelerate screening with Generation 2; key decisions still use Generation 1 plus experiment
Risks they care aboutReproducibility, reviewer scepticismOut-of-distribution errors leading to design mistakes; certification
LicensingVery important; budgets are limitedSecondary, but they want to buy fewer seats

C.2 Generation 2 models: open or closed?

"Open source" comes in four tiers, and each company's best-performing production version is often not open. All licences were checked on GitHub / Hugging Face.

TierModelLicenceCommercial use?
1. Fully openMACE code + MACE-MP-0 series (MP-0a, MP-0b3, MPA-0)MITYes
MatterSim, MatterGen (Microsoft)MITYes
Orb (Orbital)Apache 2.0Yes
Aurora / Aurora 1.5 (Microsoft; 1.5 open-sourced Jul 2026)MITYes
BioEmu (Microsoft)MITYes
neuraloperator (FNO library)MITYes
PhysicsNeMo (NVIDIA)Apache 2.0Yes
NeuralGCM (Google Research)Code Apache 2.0Yes (code)
ECMWF AIFSWeights CC BY 4.0Yes
Skala (Microsoft's deep-learning DFT functional)MITYes
2. Open weights, by requestUMA (Meta; fairchem code MIT)Weights require agreeing to the FAIR Chemistry License; commercial use allowed with geographic and use restrictionsYes (conditional)
3. Open but non-commercialGraphCast, GenCast (DeepMind)Code Apache 2.0, weights CC BY-NC-SA 4.0No
Pangu-Weather (Huawei), FuXi (Fudan University)Weights CC BY-NC-SA 4.0No
MACE-OFF23, MACE-OMAT-0, MACE-OMOL-0, MACE-MATPES, MACE-MHAcademic Software License (ASL)No
4. ClosedWeatherNext 2 / 3 (Google's production version)No weights released; data via BigQuery / Earth Engine, early access on Vertex AIPaid
GNoMEOnly predicted-structure data released; model not open—
Silurian GFT, Tomorrow.io; Ansys SimAI, PhysicsX, Neural Concept, Luminary; CuspAI, etc.Commercial servicesPaid

Patterns:

  1. Within the same company, the research version is open and the production version is closed. Google's GraphCast and GenCast have open weights but ban commercial use; WeatherNext 2/3, which run in production, are sold only as data and services.
  2. Open-source strategy mirrors the business model. Microsoft is the most open (monetising through Azure and Foundry); NVIDIA open-sources (monetising through GPUs); Meta is open but gated (no cloud business to protect); Google opens research versions and closes production ones; startups stay closed (the model is the product); Orbital is the exception (open source for influence, money from industrial hardware).
  3. Licences matter more to companies than to universities. Universities are basically unconstrained; commercial users can only pick tiers 1 and 2, or buy tier-4 services. Within the same MACE family, the early MP-0 series is commercially usable, while newer models trained on larger data such as OMAT/OMOL switched to an academic licence. In weather, AIFS (CC BY 4.0) and Aurora (MIT) are among the few options combining "operational-grade accuracy" with commercial use.

C.3 Market size: Generation 1 about $12–15B a year; Generation 2 actual revenue about $1–2B

Generation 1

SegmentAnnual sizeNotes
Engineering simulation (CAE)About $12–13B (2025)Top five hold 50–55%; consistent with a bottom-up tally by company (Ansys about $3B, MathWorks about $1.5B, plus Siemens, Dassault and Cadence)
"Simulation software", broad definition$13–27BDefinitions vary widely across research firms; some include discrete-event and training simulation
Quantum-chemistry softwareAbout $600M (2024)About 1/20 of engineering simulation
Molecular simulation / computational drug discoverySmallThe largest pure-play company, Schrödinger, has software revenue of about $200M
Weather, particle physics, molecular dynamics, robotics simulationSoftware revenue close to zeroFunded by governments and international organisations (ECMWF, CERN) or open source; the money goes to supercomputers and GPUs

Generation 2

DefinitionFigureCredibility
"Physics AI engineering simulation" market$15B (2025)Not credible: counts traditional CAE with added AI features
Multiphysics surrogate AI$1.4B (2024)Narrower definition
AI weather modelling$200M–1.1B (2025)Estimates differ by 5×
AI materials discovery$700M–1.3B (2025)Includes platforms and services
Bottom-up estimate (our judgement)Pure Generation 2 product revenue within about $1–2BStartup revenue is undisclosed, estimated at tens of millions of dollars each; big tech does not break it out

For comparison: venture capital flowing into Generation 2 companies in 2025–26 totals about $1.5B or more (CuspAI $450M, PhysicsX about $450M, Tomorrow.io $210M, Genesis AI $105M, Neural Concept $100M, Luminary $72M, Orbital $50M…). Annual funding is roughly as large as the whole segment's annual revenue: capital is racing ahead and revenue has not caught up.

C.4 sim2real accuracy: Generation 1 vs. Generation 2

(L1 = simulation vs. real experiment; L2 = AI model vs. the simulator it learned from)

FieldGen 1 vs. reality (L1)Gen 2 vs. Gen 1 (L2)Gen 2 vs. reality
Materials DFTFormation energy about 0.1–0.15 eV/atom; band gap PBE MAE >1 eV (about 50%), HSE/mBJ about 0.3 eVAbout 20 meV/atom, stability F1 >0.9Essentially the same as DFT; inherits DFT's errors
Small-molecule quantum chemistryCCSD(T) is the gold standard; common functionals have noticeably larger errorsDepends on the level of theory of the training dataCan approach chemical accuracy when trained on high-accuracy data
Binding free energy (FEP)Public 1.73 kcal/mol, private pharma 2.44 kcal/molBioEmu vs. MD <1 kcal/molSame order of magnitude as MD
WeatherOperational NWP is the baselineGraphCast beats HRES on >90% of metrics; GenCast beats ENS on >96%The exception: Gen 2 beats Gen 1 (it learns from reanalysis that fuses observations), but error against weather stations is 15–45% higher than against reanalysis
CFDIncrements usable at cruise with attached flow; wide scatter for separated flow and off-designIn-distribution relative L2 error of a few percentTwo layers of error stack
Particle physics (Geant4)Mean calorimeter energy response matches test-beam data to within a few percentGenerative models agree with Geant4 on common physics quantitiesClose to Geant4
RoboticsPure physics-simulation evaluation correlates with real robots at about 0.4—Video world-model evaluation correlates at 0.9+

The pattern: Generation 2's accuracy ceiling is Generation 1, with weather as the only exception. To genuinely reduce sim2real error, you have to change Generation 1 itself.

C.5 Recent methods for reducing the Generation 1 gap

Two main lines: use AI to correct physical approximations, and learn directly from observations and experiments.

  1. Use AI to make higher-level theory cheap.
    • Skala (Microsoft, arXiv 2506.14665, v6 Apr 2026): a deep-learned exchange-correlation functional that keeps the cost of semi-local functionals while reaching chemical accuracy on small-molecule atomisation energies, with a GMTKN55 error of 2.8 kcal/mol, better than mainstream hybrid functionals. It currently covers main-group molecules; transition metals and periodic systems are planned. MIT open source.
    • DM21 (DeepMind, Science 2021): a neural-network functional trained on fractional-charge systems to correct systematic errors such as charge delocalisation (its performance on fractional-charge systems has been questioned).
    • Neural-network wavefunctions: FermiNet and Psiformer solve the many-electron Schrödinger equation directly from first principles.
    • Better training data: OMol25 uses ωB97M-V/def2-TZVPD, clearly above the PBE level customary in materials science.
  2. Skip the simulator and learn from observations directly.
    • Aardvark Weather (Nature 2025): the first end-to-end, purely data-driven weather system, producing forecasts straight from raw observations without relying on NWP. It uses an order of magnitude fewer observations and orders of magnitude less compute, and achieves lower RMSE than the operational NWP baseline on several variables.
    • WeatherNext 3 (Aug 2026): assimilates raw satellite observations directly.
    • DiffTRe (Nature Communications 2021): trains neural-network potentials directly on experimental observables (thermodynamic, structural and mechanical properties) rather than only matching DFT.
  3. Physics + ML hybrids that replace only the least accurate piece.
    • Turbulence: ML learns RANS closure terms (tensor-basis neural networks, TBNN, and others); ML-trained LES wall models already match reference data on real aircraft configurations.
    • Climate: NeuralGCM keeps the physical dynamical core and replaces parameterisations of clouds, convection and so on with learned modules.
    • Molecular force fields: ML force fields such as MACE-OFF replace fixed-charge classical force fields.
  4. Calibration and digital twins. Robotics uses system identification, domain randomisation and 3D Gaussian splatting for real-to-sim; particle physics tunes Geant4 models with test-beam data; weather and engineering continuously inject observations through data assimilation.
  5. Higher resolution and more compute. The EU's Destination Earth climate digital twin is the first to produce operational global multi-decadal projections at 5–10 km; the extremes twin runs at 4.4 km globally and 500–750 m regionally over Europe, entering continuous operation in summer 2026.

The trend: the boundary between Generations 1 and 2 is disappearing. Skala is still essentially a DFT functional (Generation 1), yet it was trained by AI; NeuralGCM is a Generation 1 physics core plus Generation 2 learned modules. The direction is shifting from "AI replaces physics" to "AI corrects physics, then calibrates against real data".

References

Virtual cells

Proteins / RNA / genomes

Molecular dynamics / free energy

Materials / chemistry

Weather / fluids / physics / cosmology

Fusion / neuroscience / medicine / robotics / mathematics

RL for Science: industry

Appendix A: industry landscape

Appendix B: Gen-2 / Gen-3 market

Appendix C

Compiled from several rounds of web research. Most private-market valuations are media reports or figures under negotiation, not official disclosures. Data as of Sep 21, 2026.

Adapted from the Spark Science research notes (originally published in Chinese at jimmyresearch.com). Last updated Sep 21, 2026.