Abstract. We present DUAL-X, a closed-loop optimization framework that integrates interpretable, human-centric rationale extraction with gradient-based attribution to give MLFF (machine-learned force field) predictions a SHAP-like audit trail — connecting atomic-position contribution scores back to chemically meaningful descriptors (SOAP, SNAP, environment descriptors).

Why it matters. The AAAI 2026 XAI4Science track spotlights methods that open the "black box" for scientific ML; DUAL-X is the first framework to ship a practical, end-to-end XAI recipe for materials MD models.

Abstract. AutoMat is a benchmark for evaluating LLM agents on scientific reproducibility in computational materials science. Agents receive the artifacts of a published paper and must reproduce the reported result; we score both fidelity and faithfulness to the original workflow.

Why it matters. The first benchmark to put LLM "reproducibility" under the same microscope for materials science that we already use for ML benchmarks.

Abstract. ARIA (Adaptive Reasoning with Interpreted Atoms) is a causal-aware framework for rescuing LLM reasoning in trustworthy materials discovery. It combines rigid knowledge-graph retrieval, a contextual-tower contextualization module, and a tiered fallback hierarchy (Direct Match → Analogical → Fallback) to keep LLM suggestions chemically grounded.

Why it matters. Born out of the 2025 LLM Hackathon Visionary Award, ARIA is the first framework to combine KG retrieval with causal-aware hierarchical reasoning for materials science.

Abstract. We use migration pathways as diagnostic probes to benchmark specialist vs. generalist machine-learned force fields (MLFFs), exposing where task-specific fine-tuning helps or hurts generalization across chemistries.

Why it matters. Spotlight talk & travel grant at NeurIPS AI4Mat 2025; a physically meaningful, generalizable probe for MLFF generalization in materials science.

Abstract. We identify low-energy pathways by which point defects in the CsPbBr3 perovskite self-heal, pointing to design rules for more robust lead-halide perovskites.

Why it matters. Connects an atomistic self-healing mechanism to rational material design for stable perovskite devices.

Summary. A two-mode intercalation mechanism functions as an atomic switch, reversibly modulating the interlayer structure and properties of 2D materials.

Why it matters. Offers a controllable, switchable route to tune 2D material behavior for device applications.

Summary. Organizes AI-powered de novo peptide design around the end-to-end discovery workflow rather than individual model families.

Why it matters. Gives practitioners a map from computational design choices to downstream experimental validation.

Summary. A review of strategies and recent advances in engineering nanoparticles for brain-targeted drug delivery.

Why it matters. Surveys the landscape of delivery approaches to inform nanoparticle design for neurological disease.