Physics-grounded generative AI for chemistry
Designing the refrigerants the world can’t find.
Refrigerants drive ~2% of global greenhouse-gas emissions — about the same as aviation — yet after decades of screening the industry has only ~300 viable molecules. Refgen designs brand-new ones.
The problem
No molecule satisfies all the constraints
A warming planet drives cooling demand; cooling systems leak refrigerants; refrigerants are among the most potent greenhouse gases. One kilogram of R-410A warms the planet like ~2,000 kg of CO₂ — and the molecules that would replace it largely do not exist yet.
Five constraints, none solved simultaneously
- COP ≥ R-410A efficiency
- GWP < 150 regulation
- Non-flammable safety
- Non-PFAS EU REACH
- Synthesizable & stable manufacturing
Sources: McLinden et al., Nature Communications (2017); Goldszal et al., arXiv:2509.19588 (2025).
The phase-out is already in law
Every major jurisdiction has a binding HFC phase-down on the books. The market for compliant molecules is being pulled into existence by regulation.
-
2027
EU F-Gas
< 150 GWP, stationary < 12 kW
-
2029
Canada & US
HFC −70%
-
2034
Canada & US
HFC −80%
-
2036
Kigali
HFC −85%
Sources: EU Regulation 2024/573; Canada ODSHAR SOR/2016-137; US EPA AIM Act (40 CFR 84); Kigali Amendment.
The technology
Physically grounded generative AI
Refgen couples a generative model of chemistry with a rigorous thermodynamic engine in the training loop. The model doesn’t just propose plausible molecules — it is rewarded for proposing ones the physics says will actually work.
- 01
Generate
A sequence model trained on ~40M molecular structures proposes brand-new, valid refrigerant chemistries as SMILES.
- 02
Ground in physics
Each candidate is scored by physics, not guesswork: Peng–Robinson EOS, NASA polynomials and a full vapor-compression-cycle simulation predict COP, GWP, flammability and stability.
- 03
Optimize
Reinforcement learning steers generation toward the high-COP, low-GWP region — discovering molecules no screen would ever reach.
Closed-loop discovery. Experimental data from each synthesis round feeds back into the model, sharpening every next generation — and the same engine treats the vapor-compression cycle itself as a variable, tailoring molecules to a specific equipment class.
Based on the paper Refgen: De Novo Discovery of Sustainable Refrigerants (Goldszal, Calanzone, Taboga, Bacon) — arXiv:2509.19588.
Team
-
Vincent Taboga, Ph.D.
Project lead
Get in touch
Let’s design the molecule your roadmap is missing.
We work with fluorochemical producers and equipment makers facing the HFC phase-down and PFAS exposure.