Speaker
Description
Spacecraft thermal analysis traditionally relies on full scale physical simulations performed with industry standard tools such as ESATAN-TMS and Thermica. While essential for mission verification, these simulations can be computationally intensive, creating a major bottleneck for large-scale sensitivity analyses, design space exploration, and statistical Monte Carlo uncertainty quantification.
To address this limitation, OBO Space developed PIN-TAN (Physics-Informed Neural – Thermal Analysis Network), an end-to-end surrogate modelling framework, enabling real-time transient temperature predictions by combining deep neural networks with governing thermal physics equations.
The framework directly interfaces with ESATAN-TMS through a dedicated parser that ingests geometry files (.erg) and input decks (.d) to extract nodal properties (diffusive, arithmetic, and boundary), linear and radiative conductors, absorbed environmental fluxes and parametric design variables defined in GMM and TMM such as thermo-optical properties, material conductivities, contact conductances etc. A Latin Hypercube Sampling (LHS) engine explores the user-defined parameter space, orchestrates batch executions of the GMM and TMM solvers, and compiles the resulting temperature results into structured training datasets. The core surrogate relies on a Multi-Layer Perceptron (MLP) trained to predict multi-node temperatures as a function of time and operational parameters. The optimization is governed by a composite loss function that combines data regression with a physics-informed residual enforcing the nodal energy balance differential equation across conductive paths, non-linear radiative couplings, environmental fluxes, power dissipations and nodal thermal capacities. The complete workflow is packaged into an interactive dashboard, allowing engineers to monitor dataset creation, model training, inspect node-level validation, and perform real-time transient predictions and Monte Carlo sensitivity analyses.
Preliminary evaluations demonstrate that PIN-TAN achieves high physical consistency and thermal accuracy compatible with space engineering requirements for reduced models, while cutting evaluation times from traditional solver timescales down to sub-second responses. PIN-TAN provides space thermal engineers with a practical tool for rapid trade studies, thermal digital twins, and uncertainty quantification directly compatible with standard industrial workflows.