Document Type

Publication - Article

Towards Physics-Informed Neural Networks for Magma-Chamber Cooling: A Case Study of the Rio Pisco Pluton

Department

Computing

Date of Activity

Summer 2026

Abstract

Magmatic–hydrothermal systems transport heat through coupled conduction and buoyancy-driven fluid flow in porous rock, behavior conventionally modeled with grid-based finite-difference simulators such as HYDROTHERM. We demonstrate that a physics-informed neural network (PINN), built on the NVIDIA PhysicsNeMo framework using automatic differentiation and mesh-free collocation, can produce a stable two-dimensional time-dependent solution for a magma-chamber configuration based on the Rio Pisco pluton in the Peruvian Coastal Batholith. Boundary conditions and material parameters are taken from a prior HYDROTHERM study of the same pluton, and 28 temperature samples digitized from that study are used as a supervised constraint. The PINN couples Fourier conduction, advective heat transport, Darcy flow with a temperature-dependent permeability law, and a mass-conservation formulation; the mass-conservation equation is written in two-phase form, but in the regime studied here, the simulation remains below the boiling curve, so the steam-phase saturation stays at zero and the formulation reduces to its single-phase liquid–water limit. The network reproduces the conductive temperature gradient and a directionally consistent buoyancy-driven flow field, with weaker and less organized circulation than the reference simulation, and a cooling time of approximately 1.6×105 years, comparable to the ∼175,000 years reported for the matching 𝑘=10−16m2 HYDROTHERM reference scenario from which the supervised training data was digitized. We discuss the conditions under which the mesh-free, automatically differentiable PINN approach offers a useful alternative to grid-based solvers.

Link to Full Text

Share

COinS