Deep Probabilistic Surrogate Networks for Universal Simulator Approximation Andreas Munk Adam ´ Scibior Dept. of Computer Science University of British Columbia Vancouver, B.C. Canada amunk@cs.ubc.ca Dept. of Computer Science University of British Columbia Vancouver, B.C. Canada ascibior@cs.ubc.ca Atılım G¨ une¸ s Baydin Andrew Stewart Dept. of Engineering Science University of Oxford Oxford, OX2 6PN United Kingdom gunes@robots.ox.ac.uk Convergent Manufacturing Technologies Inc. Vancouver, B.C. Canada andrew.stewart@convergent.ca Goran Fernlund Anoush Poursartip Frank Wood Dept. of Materials Engineering University of British Columbia Convergent Manufacturing Technologies Inc. University of British Columbia Vancouver, B.C. Canada goran.fernlund@convergent.ca Dept. of Materials Engineering University of British Columbia Convergent Manufacturing Technologies Inc. University of British Columbia Vancouver, B.C. Canada anoush.poursartip@ubc.ca Dept. of Computer Science University of British Columbia Vancouver, B.C. Canada fwood@cs.ubc.ca Abstract We present a framework for automatically structuring and training fast, approximate, deep neural surrogates of existing stochas- tic simulators. Unlike traditional approaches to surrogate modeling, our surrogates retain the interpretable structure of the reference simulators. The particular way we achieve this allows us to replace the reference sim- ulator with the surrogate when undertaking amortized inference in the probabilistic pro- gramming sense. The fidelity and speed of our surrogates allow for not only faster “for- ward” stochastic simulation but also for ac- curate and substantially faster inference. We support these claims via experiments that in- Preprint. Work in progress volve a commercial composite-materials cur- ing simulator. Employing our surrogate mod- eling technique makes inference an order of magnitude faster, opening up the possibility of doing simulator-based, non-invasive, just- in-time parts quality testing; in this case in- ferring safety-critical latent internal temper- ature profiles of composite materials under- going curing from surface temperature profile measurements. 1 Introduction Simulators are accurate generative models that encode the knowledge of domain experts. Whether in aero- nautical engineering (Wu et al., 2018), nonlinear flow physics (Veldman et al., 2007), or stochastic generative modeling (Heinecke et al., 2014; Endeve et al., 2012; Raberto et al., 2001; Perdikaris et al., 2016), they play arXiv:1910.11950v1 [cs.LG] 25 Oct 2019