Scientific ML Agents Real-world data

Machine learning grounded in physics

Agents in flow, structure from spectra, forecasting when sensors are thin. Models and pipelines built to hold up in the real world.

Who am I? See selected work Collaborate

Hello

I’m Josef Berman - data scientist and researcher at the overlap of RL, computational physics, and scientific ML: learning in dynamic physical environments, structure and signals from scientific data, and forecasting when real-world measurements are sparse. I optimize for systems that respect physics and instrumentation, not just offline metrics.

Core & ML

Python PyTorch TensorFlow scikit-learn

Agents & RL

Stable-Baselines3 LangGraph LangChain Gradio

Scientific

PhiFlow RDKit

Selected work

FluxSwarm

RL swarms in deforming flow

Train micro-agent collectives in time-varying velocity fields - channels, vasculature, and other flows where drag and coordination matter.

GitHub

NMR2structure

1D NMR → substructures

Gradient boosting + cheminformatics to propose subfragments from 1D spectra - faster elucidation with interpretable hints.

GitHub Paper

syN-BEATS

Sparse air-quality forecasting

Neural ensemble on N-BEATS for pollutant levels when monitoring networks are thin or uneven.

GitHub Paper

EnergeticGraph

RAG + models for energetic materials

Retrieval-augmented Q&A with property predictors and graph reasoning over energetic compounds - SMILES, docs, and learned signals together.

GitHub

Let’s talk

Open to research collaborations and applied ML at the physics - chemistry - ML boundary.

Send Email