Explainable & Physics-Informed AI

Combining explainability and physical constraints in deep learning — from GeoPriorSubsNet v3 to lunar EM mapping — to build geohazard models that are accurate, interpretable, and physically consistent.

Data-Driven Learning

Multi-resolution attention fusion and encoder-decoder backbones that learn complex spatiotemporal patterns from InSAR, piezometric, and geological datasets.

Physics Heads

Dedicated physics head in GeoPriorSubsNet v3 recovers latent hydrogeological parameters — effective conductivity, porosity, elastic modulus — without extra supervision, enforcing physical plausibility at output time.

Uncertainty-Aware Evaluation

GeoPriorSubsNet's probabilistic outputs are audited with the k-diagram toolkit — ensuring that physical plausibility and statistical reliability go hand in hand across all quantile levels.

Are physics-constrained forecasts better calibrated than data-only baselines?
Do physics heads reduce physically implausible predictions?
Where do data-physics trade-offs emerge in multi-city transfer?

Example polar plots from the k-diagram toolkit

GeoPriorSubsNet v3: A Case Study in Physics-Guided AI

GeoPriorSubsNet v3 unifies a probabilistic data head (quantile outputs for cumulative subsidence and hydraulic head) with a physics head that recovers effective hydrogeological closure fields — conductivity, porosity, and elastic modulus — as latent variables. Trained on multi-city InSAR and piezometric data, it outperforms purely data-driven baselines and transfers across cities with minimal retraining.

The physics head acts as an implicit regularizer — it forces the model to learn representations that remain physically meaningful even under distribution shift across cities.

GeoPriorSubsNet v3 architecture: probabilistic data head + physics head recovering hydrogeological parameters

Frontier Application

Lunar Titanium Mapping — EM + PINNs Beyond Earth

Physics-informed neural networks extend beyond terrestrial geohazards. As part of an international expert panel (NASA / JAXA / CNSA collaboration, 2025), I applied Maxwell-equation-constrained PINNs to multi-spectral electromagnetic (MXS) data from lunar surface surveys to map titanium-rich basalt distributions — a setting where physical constraints substitute for the dense ground-truth labels that are impossible to collect on the Moon.

The Lunar Surface Electromagnetics Experiment (LuSEE)

LuSEE — Lunar Surface Electromagnetics Experiment: the EM environment targeted by PINN-based titanium mapping

From Research to Practice

GeoPrior-3.0 Forecaster brings GeoPriorSubsNet to practitioners who don't write code — a desktop GUI that covers the full pipeline from data intake and physics-constraint configuration to training, transfer learning, and operational forecasting.

GeoPrior-3.0 Forecaster desktop application — full physics-guided pipeline

Impact

Key Research Outcomes

Publications

Papers in JOSS, IEEE TPAMI, JEM, and Nature Communications — spanning explainability, physics-informed AI, and uncertainty diagnostics.

DL Frameworks

XTFT (uncertainty-rich temporal fusion transformer) and GeoPriorSubsNet (physics + probabilistic heads) — novel architectures for geohazard forecasting.

Open-Source Tools

GeoPrior-v3 and k-diagram released as installable Python packages with full documentation and reproducible tutorials.

Lunar Study

PINN-based titanium mapping for the LuSEE lunar electromagnetics experiment — extending geophysical AI to extraterrestrial environments.

Open-Source