Built by researchers who got tired of synthesizing dead ends.
Matforgelab started in 2025 in Richmond, VA, from a specific observation: R&D teams working on battery materials, specialty polymers, and structural alloys have access to capable ML models, but connecting those models to an actual synthesis workflow requires months of custom engineering that most labs simply cannot staff. The prediction tool ends up as a research prototype, not a working part of the discovery cycle.
The observation led to a clear question: what if property prediction, Bayesian candidate search, and characterization ingestion all lived in one tool designed around how bench scientists actually work, rather than around how ML engineers prefer to build? Not another Python library to integrate. An actual workflow.
That is what we are building here. We are early, we are direct with R&D teams about what the model can and cannot do, and we iterate fast based on what labs tell us is blocking them.
Alan Krishnamurthy
CEO and Founder
Alan's background spans materials science and applied machine learning, with several years spent on property prediction and synthesis planning problems at research institutions. The recurring observation was that materials informatics stalled not because the models were wrong, but because connecting them to an experimental workflow was a bespoke engineering project every time. Most lab groups could not afford to run that project in parallel with actual research.
He started Matforgelab in 2025 to build the tool that should already exist: one that handles composition-property prediction, active formulation search, and characterization reconciliation together, so R&D scientists can focus on the science rather than the plumbing.
How we work.
Every prediction ships with a confidence interval. We document where the model underperforms. A materials scientist's trust is not won by hiding uncertainty, it is won by reporting it accurately.
Features are designed around how a bench scientist runs a formulation campaign, not around what is easy to build in ML. The primary user is the R&D chemist, not the IT administrator or the ML engineer.
Formulation data, characterization uploads, and experimental records are isolated per account. Models fine-tuned on your data do not leave your account. No cross-customer pooling without explicit opt-in.