Papers, preprints, datasets, and software across machine learning, materials science, and physics.
39 works
Explainable AI applied to quantum size effects — letting domain scientists actually interrogate the model rather than trust a black-box prediction.
Defines the deletion floor for materials machine unlearning, showing why post-deletion accuracy must be evaluated against retraining, prediction change, and retained model utility.
Local motif-guided discovery of hydrogen-stabilizing environments in refractory high-entropy alloys.
A model-free reference for separating visual perception errors from reasoning errors in crystal-structure benchmarks.
How far can graph generative models extrapolate beyond the size and complexity of their training set? We characterize the extrapolation frontier across multiple architectures.
Clifford algebra equivariant interatomic potentials for heterogeneous catalysis.
Explainable ML over high-throughput electronic-structure calculations for molecular H₂ adsorption on single-atom-doped TiO₂ nanoparticles — 30 dopants across perpendicular and parallel geometries. A three-stage feature selection plus symbolic regression distills an interpretable model governed by H–H activation, dopant–H₂ distance, and π-electron rearrangement, yielding a mechanism-based screening rule that flags Group 4 and 5 transition metals as the most robust dopants for reversible hydrogen storage.
A descriptor-guided thermodynamic screen of H₂ adsorption on single-atom-doped anatase TiO₂ nanoparticles — DFTB + conceptual DFT descriptors + symbolic regression with leave-one-out cross-validation. Identifies a compact electrophilicity-family expression as the most generalizing structure–property law in a small-data regime, and flags Nb and Zr as the best-balanced dopants for reversible hydrogen storage.
Engineered MgO nanoparticles with electronic signatures tuned for energy applications, with ML guiding the structure-property mapping.
A benchmark that quantifies structural hallucination, consistency, and reasoning gaps in foundation models when they try to do materials science.
A benchmark for learning scale-dependent geometric invariances in 3D materials — pushing GNNs and equivariant models past the small-system regime where most papers live.
An equivariant convolutional network that predicts ground-state eigenvalues of materials directly from electronic shell structures. Outperforms DFT for the same regime at a fraction of the compute, and beats modern neural-wavefunction baselines like PsiFormer and FermiNet on our benchmark.
A multimodal benchmark for visual learning of atomic interactions. Treats the question of whether foundation models can reason about atomic-scale geometry the way they reason about photographs.
A multiphase, temperature-resolved, multimodal dataset for crystalline materials — the kind of training data that lets vision-style models actually generalize across phase boundaries instead of memorizing one regime.
A transformer text encoder fused with physics-informed losses to make quantum-chemistry workflows accessible without specialist tooling.
A human-in-the-loop framework for richer atomic-geometry representations in materials science — moving past the standard graph-of-atoms abstraction.
An agentic AI system for explainable quantum chemistry — tools and reasoning glued together so a model can answer chemistry questions with reproducible steps rather than confident guesses.
Computer-vision-based prediction of hydrogen adsorption in copper nanoclusters that needs orders of magnitude less data than the chemistry-feature baseline.
A study of when (and where) molecular graph neural networks actually work in materials science — and where the seemingly obvious inductive biases quietly fail.
A multi-modal, multi-task benchmark for temperature-dependent crystalline materials. Built to stress-test models on the kind of phase-aware reasoning that real materials work demands but most benchmarks ignore.
How well do multimodal foundation models reason about crystals? We stress-test them on tasks that require real crystallographic understanding rather than statistical mimicry.
A multimodal neural network that predicts nanoparticle properties from raw simulation outputs, learning across modalities that usually demand bespoke pipelines.
An integrated ML + DFT study of carbon doping in TiO2 nanoparticles and what that does to their electronic properties.
A transformer-based focus-detection method for wide-field interferometric microscopy, fast enough to run in the imaging loop instead of offline.
Deep-learning-based focus positioning for laser systems hitting rough surfaces, where classical autofocus stumbles.
Noise-robust focal-distance detection in laser material processing, combining CNNs with Gaussian process modeling for honesty about uncertainty.
High-precision, real-time focus detection for laser material processing — the version that pushed me to look at this whole class of problem more carefully.
A robust focus-detection pipeline for laser micromanaging that holds up under measurement noise — the deployable counterpart to the Optics journal work, presented at CLEO.
No matching publications. Try another search or filter.