AI for scientific discovery

Reinforcement Learning via Symbolic Feedback (RLSF) – Chemistry

TL;DR:
RLSF introduces a new way to fine-tune Large Language Models (LLMs) for molecular design and synthesis by combining reinforcement learning with token-level symbolic feedback from cheminformatics tools such as RDKit Instead of a single success/failure score, RDKit provides fine-grained chemical error signals (e.g., valence violations, missing functional groups, or conservation-law breaks) for every token in a generated SMILES string. These signals drive a Proximal Policy Optimization (PPO) loop to iteratively improve the LLM.

Key Chemistry Tasks and Gains:

Molecule generation: +8 – 14 % exact match and up to +58 % validity over supervised fine-tuning; outperforms GPT-4 despite using ≈1000× fewer parameters.
Forward synthesis: +12 % exact match and higher product validity by enforcing conservation of atoms and correct reaction syntax.
Retrosynthesis: +12 - 34 % exact match improvements by leveraging RDKit to back-check precursors and penalize chemically impossible suggestions.

Scientific Impact and Next Steps
RLSF demonstrates that symbolically guided reinforcement learning can dramatically boost the accuracy and chemical validity of small open-source models.
The project team is extending these techniques to material science and discovering physics theories, using domain-specific symbolic engines to drive discovery in crystal design, band-gap prediction, and neutrino mass theories.

Discovering Laws of Physics via Interpretable Siamese Neural Networks

TL;DR:
They develop interpretable Siamese Neural Networks that detect similarity among data points in theoretical physics (e.g. events in special relativity, electromagnetic field transformations, particle motion in central potentials). In training to cluster similar instances, the model also learns symmetry invariants and conserved quantities without prior domain knowledge.

Logic Guided Genetic Algorithms (LGGA)

TL;DR:
LGGA augments symbolic regression (SR) with auxiliary truths (domain‐specific known facts) to guide equation discovery. It integrates these truths into scoring (loss) functions and data augmentation to make SR more data efficient. Compared to standard SR tools, LGGA can improve data efficiency by up to ~62% in experiments.

Logic Guided Machine Learning (LGML)

TL;DR:
LGML is a two‐phase framework combining a learning model (to propose symbolic expressions from data) and a logic solver (to verify consistency of these expressions against known auxiliary truths). When the logic phase finds inconsistencies, it provides counterexamples back to the learning phase, in a feedback loop. It can learn expressions for things like the Pythagorean theorem and sine function, with many orders of magnitude better data efficiency than standard neural network approaches.