PhD Candidate in Physics · University of Florida · Gainesville, FL
I am a final-year PhD candidate at the University of Florida, advised by Richard G. Hennig and Peter J. Hirschfeld. I work on generative AI for materials discovery. I build diffusion, flow-matching and stochastic-interpolant models that propose new crystal structures, and I use reinforcement learning to steer them toward materials that are stable and have a target property. Machine-learned potentials and DFT then check the candidates before the most promising ones go to experimental collaborators for synthesis.
Much of my work so far is on superconductors. In summer 2026 I was a research intern at the Center for Nanophase Materials Sciences at Oak Ridge National Laboratory, where I worked on inverse design of 2D defect structures from a target density of states and on restoring beam-damaged electron microscopy images.
I expect to defend my thesis in December 2026 and am looking for research positions starting in 2027.
npj Computational Materials 12, 286 (2026)Spotlight, NeurIPS 2025 AI4Mat
We fine-tune a crystal-structure foundation model with an adapter and classifier-free guidance so that it generates structures with a target superconducting Tc. Of 200,000 generated structures, 773 passed ML and DFT screening with Tc above 5 K.
arXiv:2610.03880 (2026), under review. Presented at the MRS Spring Meeting 2026.
We fine-tune the OMatG generator with group-relative policy optimization, with the policy gradient acting directly on its discrete atom-type transitions. Our reward raises the yield of metastable, unique and novel structures from 13.4% to 45.5% on a community benchmark. We also show that reinforcing atom types invites reward exploitation unless the reward has explicit guards, and that single-element structures inflate the benchmark metric without giving any new compounds.
Computational Materials Science (accepted). Earlier version at the NeurIPS 2023 ML4PS workshop.
We benchmark the UF3 potential against GAP, MTP, SNAP and NNP, and test how well it transfers by predicting melting points from a model trained only on solid-state data.
npj Computational Materials 12, 95 (2026)
An ensemble of equivariant GNNs (BEE-NET) predicts the Eliashberg spectral function and Tc, and drives a screening pipeline that ends in DFT. Two of the predicted compounds were synthesized and confirmed to superconduct.
Journal of Chemical Information and Modeling 66(15), 8860–8874 (2026). Earlier version at the ICML 2025 GenBio workshop.
A comparison of guidance methods, including classifier-free guidance, autoguidance and model guidance, for property-conditioned molecule generation with flow matching.
International Conference on Machine Learning (ICML), PMLR 267 (2025)
OMatG is a stochastic-interpolant framework for crystal structure prediction and de novo generation, with discrete flow matching for atomic species.
Journal of Chemical Theory and Computation 21(8), 3913–3921 (2025)
An equivariant graph network trained on DFT spin-switching energies for 1,439 materials, combined with a relevance-based classifier, finds spin-crossover candidates about four times more often than conventional high-throughput screening.
* Equal contribution. Full list on Google Scholar.
github.com/paprakash/GuidedMatDiffusion
GuidedMatDiffusion data
Dataset released with the guided diffusion paper.
huggingface.co/datasets/paprakash
OMatGRPO
Reinforcement-learning fine-tuning of the OMatG crystal generator, with trained models on Hugging Face.
github.com/paprakash/OMatGRPO
OMatG
Stochastic-interpolant framework for crystal structure prediction and de novo generation, in PyTorch Lightning.
github.com/FERMat-ML/OMatG
BEE-NET
Equivariant GNN ensemble that predicts the Eliashberg spectral function and Tc.
github.com/henniggroup/BEE-NET
stem_infilling
Damage-aware restoration of STEM images that avoids erasing real defects.
github.com/paprakash/stem_infilling