An R&D platform for Farr Polychem that helps chemists reach the next viable formula with fewer expensive guesses.
Built for Farr Polychem, the platform turns lab history, testing data, and AI prediction into one practical R&D workflow so teams can spend less time guessing and more time validating the right formulation.
Seven models work together to improve prediction quality across different polymer properties
Inverse optimization searches for formulas that hit the target spec instead of just scoring past experiments
Materials, suppliers, test results, predictions, and optimization live in one R&D workflow
Built for production use, with training checkpoints and deployment discipline baked in
Farr Polychem's R&D team develops polymer formulations — compounds with specific physical and chemical properties for industrial applications. The traditional process is expensive trial and error: mix a formula, test it, wait days for results, adjust, repeat. With hundreds of possible ingredient combinations and dozens of target properties, even experienced chemists spend most of their time on formulations that don't work. Lab time is expensive. Materials are expensive. The feedback loop is slow.
We built an R&D platform that turns Farr's accumulated lab history into a predictive advantage. Instead of starting each formulation from scratch, chemists can now query the system: what combination of ingredients is most likely to hit these target properties? The platform answers using seven specialized models trained on Farr's own experimental data. The system has three layers: a data layer that captures materials, suppliers, test results, and experimental history in one structured workflow; a prediction layer where seven models work together to score candidate formulations across different polymer properties; and an optimization layer that runs inverse searches — given a target spec, find the formulation most likely to hit it.
Seven prediction models needed different architectures for different property types — training them on limited lab data required careful regularization and ensemble techniques
Inverse optimization searches a high-dimensional ingredient space with real-world constraints: ingredient availability, cost limits, supplier restrictions made the BoTorch acquisition function non-trivial
The platform had to feel like a tool chemists would actually use daily — showing ingredient percentages, property predictions with confidence intervals, and test-vs-predicted comparisons in formats lab teams understand
Training checkpoints and deployment discipline were essential because bad predictions don't just waste time — they waste expensive materials
Farr's R&D team now reaches viable formulations in fewer experimental cycles. The platform narrows the search space so they can focus expensive lab time on the most promising candidates. Training data grows with every experiment, so the models improve as the team uses the system. The prediction-to-validation feedback loop that used to take weeks now takes days.