Refgen: Sustainable Refrigerant Discovery

• Refgen is a generative pipeline for sustainable refrigerant discovery that integrates LLM-based SMILES generation with physics-informed property predictors.
• It combines sequence-model generation, reinforcement learning under thermodynamic constraints, and simulation to balance efficiency, safety, and environmental impact.
• Empirical tests demonstrate that Refgen replicates known trends and produces novel refrigerant candidates meeting industry standards.

Refgen is a generative pipeline for sustainable refrigerant discovery that integrates machine learning with physics-grounded inductive biases in a domain where “only about 300 refrigerants are known” and “only a few additional candidates have been suggested without experimental validation.” The framework combines a sequence-model generator, predictive models for critical properties, equations of state, thermochemical polynomials, and full vapor compression cycle simulations, then applies reinforcement learning fine-tuning under thermodynamic constraints so that generated molecules can be evaluated for efficiency, safety, and environmental impact during optimization rather than only after sampling (Goldszal et al., 23 Sep 2025).

1. Problem setting and scope

Refgen is situated in refrigerant discovery for air-conditioning systems, where “most refrigerants currently used in air-conditioning systems, such as hydrofluorocarbons, are potent greenhouse gases and are being phased down.” The paper frames the central methodological difficulty as data scarcity: “This scarcity of reliable data limits the effectiveness of purely data-driven methods.” The resulting design objective is therefore not merely valid molecular generation, but valid generation under strong physical priors that remain usable when experimental measurements are sparse (Goldszal et al., 23 Sep 2025).

The target design space is explicitly multi-objective. Refgen seeks compounds that balance “high thermodynamic efficiency, safety, and minimal environmental impact,” operationalized in the paper through properties such as coefficient of performance (COP), volumetric cooling capacity ($Q_{vol}$), critical temperature ($T_c$), global warming potential (GWP), and lower flammability limit (LFL). The framework also emphasizes “de novo refrigerant discovery beyond the known set of compounds,” rather than ranking only within a fixed candidate library (Goldszal et al., 23 Sep 2025).

2. Generative backbone and molecular representation

The generative backbone is “a LLM (Llama 3.2-1B) fine-tuned to generate syntactically valid and chemically meaningful SMILES molecular strings.” Molecular representation is therefore sequence-based from the outset, and the entire pipeline is organized around direct learning and generation over SMILES strings rather than graph generation or enumerative screening (Goldszal et al., 23 Sep 2025).

Supervised fine-tuning is performed on “curated molecular corpora (PubChem, ChEMBL, SureChEMBL)—filtered and processed to yield ~37M clean SMILES with delimiters.” The data-cleaning stage is described as enforcing molecular realism by “removing salts, invalid substructures, abnormal atom valences.” The paper also notes “data augmentations via alternative traversals,” which are used to improve generalization over molecular string representations. In the later reinforcement-learning stage, validity remains a hard constraint: the framework requires that a SMILES string be “RDKit-parseable, no radicals, valence-correct,” and “any molecule failing RDKit parsing or valence checks receives zero reward” (Goldszal et al., 23 Sep 2025).

3. Physics-grounded predictive stack

A central feature of Refgen is that the generator is coupled to “sequence models (using SMI-TED, a Transformer variant) trained end-to-end on SMILES to predict key physical and chemical properties.” The predicted quantities include thermodynamic descriptors—critical temperature ($T_c$), critical pressure ($P_c$), and acentric factor ($\omega$); engineering performance metrics—COP and $Q_{vol}$; safety and environmental indicators—GWP and LFL; and “NASA polynomials for $C_p, H, S$” (Goldszal et al., 23 Sep 2025).

The paper’s key methodological claim is that when “direct property data is scarce, models are trained on sub-properties for which data is more available, then calculate key metrics via physics-based relations/equations of state to enforce consistency.” The Peng–Robinson equation of state is embedded as

$$P = \frac{RT}{V_m - b} - \frac{a \alpha(T)}{V_m (V_m + b) + b (V_m - b)},$$

with $a$, $b$, and $T_c$0 treated as functions of predicted $T_c$1, $T_c$2, and $T_c$3. These quantities support “simulation of full vapor-compression cycles,” from which COP and volumetric cooling capacity are computed as

$T_c$4

Thermochemical consistency is introduced through NASA polynomials:

$T_c$5

$T_c$6

$T_c$7

For environmental impact, GWP is modeled through atmospheric lifetime and radiative efficiency:

$T_c$8

with

$T_c$9

The hydroxyl radical reaction rate is in turn supported by “a group-contribution method,” where “the molecule is broken into chemical groups, each with known kinetic contributions, and rates are summed” (Goldszal et al., 23 Sep 2025).

4. Reinforcement learning under thermodynamic constraints

Refgen applies “Group-Relative Policy Optimization (GRPO)” for post-training reinforcement learning. In this stage, the policy is rewarded for generating molecules that satisfy “property constraints derived from physics-informed predictors,” while still meeting hard chemical validity constraints. The reward is described as “a weighted combination of property achievement, diversity, and hard chemical validity filtering” (Goldszal et al., 23 Sep 2025).

The weighting scheme reported in the paper assigns the main weights as follows: “COP and $T_c$0 (0.4), $T_c$1 (0.4), Molecular length (0.1), GWP (0.05), LFL (0.05).” Property rewards are “soft-constrained via exponentially/gaussian-shaped rewards, plateauing in the ideal range, decaying elsewhere.” For $T_c$2, the paper gives the example

$T_c$3

To prevent collapse toward repetitive candidates, “diversity is measured by frequency counts and Tanimoto similarity; highly repetitive molecules are penalized.” The paper characterizes the resulting optimization as one in which “tradeoffs are captured automatically by the reward system,” so that the search does not maximize a single property in isolation but instead “find[s] a feasible optimum given all constraints” (Goldszal et al., 23 Sep 2025).

5. Data scarcity, extrapolation, and inductive bias

The paper repeatedly treats data scarcity as the defining constraint of refrigerant discovery: “Experimental property datasets for refrigerant-relevant thermodynamic properties are extremely limited (hundreds of molecules, not millions).” Refgen’s response is to anchor learned predictions in analytical structure rather than to rely only on direct end-to-end mapping from SMILES to final engineering targets (Goldszal et al., 23 Sep 2025).

This physics embedding is presented as the main reason the framework can move beyond interpolation over known compounds. The paper states that the models “compute downstream properties via symbolic/analytical expressions (e.g., EOS, thermodynamic cycles, group-contribution), which are valid even for out-of-domain or novel molecules—assuming the property predictors supply plausible sub-properties.” This suggests a broader extrapolation regime than direct black-box prediction over scarce labels. The reported empirical support is that “SMI-TED-based property predictors match or exceed tree-based baselines, especially for out-of-distribution CoolProp test sets (e.g., R²=0.94 for $T_c$4 on CoolProp vs XGBoost 0.55).” The same section further reports that “the group-contribution method for $T_c$5 generalizes much better to hundreds of thousands of unseen small molecules than direct neural prediction” (Goldszal et al., 23 Sep 2025).

6. Reported outcomes and discovered candidates

The paper reports both scientific reproduction of known trends and generation of new candidates. On the scientific side, Refgen “replicates known empirical trends, e.g., the famous tradeoff: as $T_c$6 increases, so does COP, but $T_c$7 decreases.” On the generative side, “RL-fine-tuned Refgen models yield distributions of generated molecules that differ significantly from the base LLM, with a shift toward chemically plausible, small, fluorinated, and/or nitrogenous structures,” including “new classes of small N–F containing molecules with promising properties not in known datasets” (Goldszal et al., 23 Sep 2025).

The paper also reports engineering-level screening behavior. “Simulated vapor-compression cycles for generated SMILES yield COP and $T_c$8 in competitive ranges with industry standards (e.g., rivals R-410A, one of the best commercial refrigerants).” When threshold filters are applied—“COP > 5, $T_c$9K, LFL > 0.1, GWP < 10”—the result is “a diverse set of molecules satisfying all, including non-PFAS candidates.” Specific examples given are C(F)N(F)N(F)F and N(F)C(F)N(F)F, which are reported to have “COP ~6.7, $P_c$0 ~338K, GWP under 4, LFL >0.5, and system sizes (represented by $P_c$1) >5, all in the ideal range” (Goldszal et al., 23 Sep 2025).

A plausible implication is that Refgen reframes refrigerant discovery as a single optimization loop in which generative modeling, property prediction, thermodynamic simulation, and constraint enforcement are tightly coupled. In the paper’s formulation, the practical contribution is not only valid molecular generation, but a way to use scarce data “effectively” by embedding the physical structure needed for de novo discovery beyond the currently known refrigerant set (Goldszal et al., 23 Sep 2025).