Skip to main content

Toward Machine Learning Optimization of Experimental Design

Atılım Güneş Baydin‚ Kyle Cranmer‚ Pablo de Castro Manzano‚ Christophe Delaere‚ Denis Derkach‚ Julien Donini‚ Tommaso Dorigo‚ Andrea Giammanco‚ Jan Kieseler‚ Lukas Layer‚ Gilles Louppe‚ Fedor Ratnikov‚ Giles C. Strong‚ Mia Tosi‚ Andrey Ustyuzhanin‚ Pietro Vischia and Hevjin Yarar

Abstract

The design of instruments that rely on the interaction of radiation with matter for their operation is a quite complex task if our goal is to achieve near optimality on some well-defined utility function U, such as the expected precision of a set of planned measurements achievable with a given amount of collected data. This complexity stems from the interplay between physical processes that are intrinsically stochastic in nature—the quantum phenomena that take place at the subnuclear level—and the vast space of possible choices for the physical characteristics of the instrument and its detection elements, as defined in its design phase. The precision of pattern recognition of detected signals and the power of information-extraction procedures that directly affect the value of U both depend on these characteristics. In the majority of realistic cases, U may be represented as a combination of performance and cost considerations that should be balanced within reasonable limitations. Neural networks are naturally suitable for the task mentioned above. They can also be effectively used as surrogates for simulators to enable gradient-based optimization in cases where a simulator is nondifferentiable. In addition, automatic differentiation (AD) techniques developed in the 1980s [2] and now commonly available in the most popular machine learning (ML) frameworks make it possible to rely on efficient implementations of the back-propagation algorithm. The MODE Collaboration (an acronym for Machine-learning Optimized Design of Experiments) aims at developing tools based on deep neural networks and modern AD techniques to implement a full modeling of all the elements of experimental design, achieving end-to-end optimization of the design of instruments via a fully differentiable pipeline capable of exploring the Pareto-optimal frontier of U. Exploratory studies have shown that very large gains in performance are potentially achievable, even for very simple apparatus. MODE has the goal to show how those techniques may be adapted to the complexity of modern and future particle detectors and experiments, while remaining applicable to a number of applications outside of that domain. Below we succinctly describe the research program of the MODE Collaboration.

Journal
Nuclear Physics News
Number
1
Pages
25–28
Publisher
Taylor & Francis
Volume
31
Year
2021