Skip to content

Projects · Ongoing · 2024–present

LiL-Q: convex training for physics-informed models

Recasts physics-informed neural network training for nonlinear PDEs as a short sequence of linear least-squares solves, removing the nonconvexity that makes standard training slow and unreliable. Tested on seven benchmark problems and on reservoir flow, against neural-network and finite-volume baselines.

Role
Lead developer, M.S. thesis and PhD research
With
Abdul-Akeem Olawoyin, Hongchao Tan, Rami M. Younis
Topics
Physics-informed ML, Quasilinearization, Least squares, Reservoir simulation
Four tall panels of the same 1,200 by 2,200 ft field. The first shows permeability, with tighter rock in the upper part. The second and third show pressure from finite volume and from LiL-Q, nearly identical, rising from 3,000 psi at the bottom to 6,000 psi in a band below the top edge. The fourth shows their difference, near zero over most of the field, with patches of up to about 280 psi of either sign in that top band.
Pressure on SPE10 Model 2, layer 3, held at 6,000 psi along the top and 3,000 psi along the bottom. LiL-Q, one least-squares solve for 3,169 coefficients, differs from the finite-volume reference by 27.3 psi RMS, mostly in tight rock below the top edge. Drawn from the public lilq-pinn results.

The problem

Physics-informed neural networks promise simulation without a mesh: a network is trained until it satisfies the governing equations at a set of points. In practice training is slow, needs careful tuning, and often fails to converge, even on linear problems. Despite years of interest, that has kept these models in proof-of-concept papers rather than in reservoir engineering workflows.

The idea

Two things make the training problem nonconvex. One is the physics: the equations of multiphase flow are nonlinear. The other is the network: its nested activation functions make the loss nonconvex even when the equation is linear.

LiL-Q removes both. Bellman–Kalaba quasilinearization replaces the nonlinear equation with a sequence of linear ones, each linearized about the current solution. A trial space that is linear in its trainable coefficients, such as spectral polynomial and trigonometric bases or random-feature networks, then turns each linear problem into a least-squares problem, solved directly by one QR factorization instead of thousands of gradient steps. Comparing all four combinations of network or linear trial space, with or without quasilinearization, showed that removing only one source of nonlinearity is not enough.

What it showed

We tested LiL-Q on seven problems: the Bratu, Burgers and Buckley–Leverett equations, Kovasznay flow, 3D unsteady Beltrami flow, linear elasticity, and Darcy flow on SPE10 permeability. It converged in single-digit outer iterations at all but the coarsest basis sizes, with runtimes up to three orders of magnitude shorter than standard training. On the Navier–Stokes benchmarks it matched published solvers with up to two orders of magnitude fewer parameters, and it recovered the elasticity solution to machine precision.

Where it stands

Learned solvers have to be judged against the classical methods engineers already use. On the SPE10 pressure problem LiL-Q took about 24 seconds where a standard physics-informed network took about 1,600, but a finite-volume solver took about a tenth of a second. Closing that gap, with bases that keep the least-squares systems sparse, and handling fully coupled pressure and saturation are the open problems our ECMOR 2026 paper sets out.

Papers and code

LiL-Q was my M.S. thesis at Texas A&M, defended in June 2026. The methods paper is on arXiv, the reservoir application is in the ECMOR 2026 proceedings, and I presented the work at AESCAPE 2026 in Dallas. The code is open source as lilq-pinn, with a guide to reproduce every result.