Home / Science
Our method
AI models map the condition space; GPU-accelerated annealing searches that map in parallel. They hand results back and forth, zooming in each round, until the optimum is found with as few experiments as possible. The order of each step adapts to the process.
AI
AI models learn from experiments and simulation data, predicting outcomes and their uncertainty across the whole window.
GPU annealing
Thousands of annealing chains run in parallel on GPUs over the AI model, escape local optima and lock onto the most promising region.
Repeat
Each round narrows the search and sharpens the model, so the next experiment is always the most informative one.
Publications
J. Cheminform. · 2025
Shows how a Digital Annealer Unit can search very large reaction-condition spaces for higher yields. Models trained on high-throughput and Reaxys data match classical ML accuracy while screening billions of conditions far faster.
Read the paper ↗Green Chem. · 2025
Uses active learning to optimize a multi-parameter conversion of chitin into the nitrogen-rich furan 3-acetamido-5-acetylfuran, outperforming trial and error guided by chemical intuition. It reaches up to 70% yield from N-acetylglucosamine, works directly on dry shrimp shells and scales to 4.5 mmol in a reusable medium without toxic high-boiling solvents.
Read the paper ↗npj Quantum Inf. · 2023
Proposes a hybrid quantum chemistry and quantum computing workflow to predict the dominant tautomeric form of drug-like molecules, a hard but important step in computer-aided drug design.
Read the paper ↗Nat. Rev. Chem. · 2022
A perspective that sets out method development and evaluation guidelines for machine-learning studies in chemistry, so new algorithms can be compared fairly.
Read the paper ↗J. Org. Chem. · 2021
Combines random sampling and active machine learning to optimize the synthesis of isomacroin while running only 3% of possible reactions, then uses kinetic modeling to confirm the global optimum.
Read the paper ↗Cell Rep. Phys. Sci. · 2020
Introduces LabMate.ML, an interpretable adaptive algorithm that optimizes many reaction parameters from a random sample of just 0.03–0.04% of the search space, validated in nine prospective studies.
Read the paper ↗Nat. Rev. Chem. · 2019
A review of how AI is changing synthetic organic chemistry, from route planning to automation, and of the problems that remain open.
Read the paper ↗
