Where AI learns to do science, and scientists learn to build with AI.
Spark Science is an open community of researchers and developers working on the layer between frontier models and real laboratories: research agents, benchmarks, verifiers, open challenges and shared data.
Rooted in the Tsinghua Spark community
Our name comes from the Tsinghua University Spark Program, a student innovation program whose motto is “a single spark can start a prairie fire.” Since 2007 it has brought together undergraduates from different departments to do research beyond the classroom.
Spark Science starts from that same habit: people from different fields working on one hard problem. AI for Science needs exactly this mix of model builders, bench scientists and engineers.
- ProgramTsinghua University Spark Program
- Founded2007
- FormatCross-disciplinary cohorts selected each year from second-year undergraduates
- TraditionFrontier talks, independent research projects, industry visits and Spark Day salons
Alumni and companies
Researchers, founders and companies from our network who are building at the intersection of AI and science.
Name
Field · Organization. One line on their work.
Name
Field · Organization. One line on their work.
Company
Domain. One line on what they build.
What we believe
Four working hypotheses that shape what we build.
Science is the next verifiable domain.
After code, research is the next field where feedback can be checked and progress can be measured.
The bottleneck sits between models and labs.
Frontier models are capable and science has hard problems. What is missing is the layer that turns lab work into something models can read, run and learn from.
Generation is outpacing verification.
Our review of recent arXiv work on AI research agents found that producing results is easy, while checking them is the real constraint.
The edge is filtering, not knowing more.
Every frontier model knows many fields. Lasting value comes from good filters, leak-proof historical backtests and cheap validation loops.
Open tools
Software for researchers who want AI in their workflow, and for developers who want to build for science.
Spark Research
A project-centric research agent. It keeps a full evidence graph of every step and connects dry and wet work, starting with lab-robot simulation.
View on GitHub →Spark Research Benchmark
Measures how well AI plus a harness can actually do research, across six research skills and four levels of autonomy. Uses delayed ground truth and leakage audits. Models never grade themselves.
Science Model API
One API for running scientific AI models on demand, with GPUs that scale up and down per job.
Open challenges with real lab readouts
Public competitions where designs are tested in automated wet-lab experiments, so the leaderboard reflects reality rather than proxy scores.
RNA Design Challenge
Fresh targets, automated validation and open results. Candidate tracks include inverse design toward a target structure, ligand-responsive switches and aptamer optimization.
Propose a challenge
Have a problem with a measurable readout? We help turn it into a fair, reproducible challenge with held-out ground truth.
Open research library
Field notes, landscapes and datasets on AI for Science, written for practitioners.
AI for Bio
Drug discovery, organoids, CROs, tools and challenges, organized into a browsable library.
Browse →Benchmark teardown
A structured breakdown of existing AI research benchmarks, plus a framework of human research skills.
Compute field tests
Hands-on comparisons of cloud GPU platforms for scientific workloads.
How we think about the problem
Open frameworks we use and welcome critique on.
Design and Execute: two tracks of a closed loop
Most AI for Science stories cover only the Design track. Real lab work needs both.
- Design decides which experiment to run. Its output is an experiment order with metrics, precision and stopping rules, not a paper.
- Execute decides whether the experiment was done right. It returns attributable results with a full record of what happened.
- Deviation flags tell Design when a data point must not be used. Without this link, the loop is statistically contaminated.
A layered stack for scientific AI
- Data access: literature, patents, databases, instruments and lab systems.
- Domain adaptation: fine-tuning and reinforcement learning on verifiable tasks.
- Verifier environments: discipline-specific checkers, evals and benchmarks.
- Workflow harness: agent orchestration, deliverables, human review and audit trails.
- Data flywheel: research trajectories, including failed experiments, fed back into training.
Data principles
- Private data stays inside its owner's environment.
- Failed and negative experiments are kept as first-class data.
- Every dataset records its source, intended use and ownership.
Domains we work on
Get involved
Whether you run experiments or write code, there is a place to contribute.
Contribute to Spark Research
Add skills, data connectors or compute backends to the open-source agent.
Open issues →Write benchmark tasks
Turn a research judgment from your field into a task with a verifiable answer, or join blind expert grading.
Enter a challenge
Compete in open challenges where results are read out in real experiments.
Request early access
Try the Science Model API and benchmark harness before public release.