A Deep Learning Surrogate Framework for High-Dimensional Regression Problems in Mechanical Engineering

Valentin Mouton, Adrien Mélot

Keywords: Deep learning, Surrogate model, High-dimensional regression, Mechanical engineering

TL;DR: A deep learning model trained on a large-scale synthetic dataset demonstrates the viability of deep learning-based surrogates for high-dimensional regression in mechanical engineering.

Abstract: This paper introduces the first large-scale deep learning-based surrogate model for high-dimensional regression tasks in real-world mechanical engineering contexts. The model, comprising 43 million parameters, is trained on a custom in-house dataset, containing 2.8 billion data points from 31 million samples that are generated entirely through easy-to-evaluate, physics-based simulations. Each sample consists of 26 scalar features and 64 scalar targets. This large-scale synthetic dataset enables the training of deep neural networks over exhaustive and realistic mechanical design spaces. It exhibits complex statistical characteristics, including zero inflation, mutually exclusive features, strong multicollinearity, and a mix of real- and integer-valued data. Despite the scale and complexity of the dataset, the model is trained using entry-level consumer-grade graphics cards, thereby demonstrating the practical viability of deep learning for regression tasks in mechanical engineering applications.

Primary Area: Machine learning for sciences (e.g. climate, health, life sciences, physics, social sciences)

Submission Number: 13213