Graduate Research Assistant, Texas A&M University (Computational Reservoir Engineering) — hybrid Newton + Physics-Informed Deep Learning training for nonlinear PDEs
Performed AI model training for physics-informed learning by developing hybrid training frameworks for solving nonlinear PDEs with PINNs and related neural operators. Implemented an inexact Newton optimization strategy that tunes loss tolerances during training to improve convergence and reduce required epochs. Built custom PyTorch training loops using automatic differentiation to enforce physics-constrained losses on multiple benchmark PDE systems. • Hybridized Bellman-Kalaba quasilinearization with PINN workflows for hyperbolic conservation laws • Compared DeepONet against fully connected networks for learning linearized update operators under varying boundary conditions • Trained surrogate models on benchmark equations including Burgers’ and Buckley–Leverett (two-phase flow) and the Brusselator reaction–diffusion system • Used L-BFGS and Adam optimizers to achieve reported reductions in training epochs vs baseline nonlinear PINNs