Use-Inspired Bayesian Optimization for Engineering Design: From Materials and 3D Printing to Circuits

Alaleh Ahmadianshalchi

doi:10.7273/000007247

The problem of adaptively selecting a sequence of experiments to optimize multiple conflicting objectives has been a challenging obstacle in many real-world scenarios. This process, often referred to as adaptive experimental design arises in many real-world domains including material discovery, analog circuit design, and additive manufacturing. For instance, in analog circuit design, one of the key challenges is optimizing circuit parameters that balance objectives efficiency, voltage, and ripple by conducting the fewest number of expensive circuit design simulations. Similarly, in biomedical engineering, optimizing a 3D printer’s parameters for creating the perfect presurgical organ models while spending the least amount of money on required materials, involves trading-off print quality, speed, and cost of material. In such multi-objective optimization (MOO) problems, we have the ability to evaluate any candidate solution by performing expensive experiments, where the cost is measured in terms of consumed resources, be it computational or physical resources (e.g., cost of materials required to perform a lab experiment). The overarching goal is to approximate the optimal Pareto set, a set of non-dominated solutions that represent the best possible trade-offs among objectives, while minimizing the total resource cost of experiments. This thesis studies new challenges in adaptive experimental design within the framework of Bayesian optimization (BO) that are directly inspired by use-cases in real-world engineering design problems. First, we need to perform expensive experiments to determine if an input design is feasible or not (aka black-box constraints); design space contains a small fraction of feasible inputs; and there are preferences over multiple conflicting objectives. Second, we need to select a batch of experiments to find high-quality and diverse Pareto front solutions. Third, we have a small experimental resource budget which requires a planning or non-myopic approach to find high-quality Pareto solutions within the resource budget. For the first challenge, we introduce a preference-aware constrained multi-objective Bayesian optimization algorithm that allows practitioners to specify preferences over objectives, ensuring that the optimization process aligns with domain-specific priorities while handling expensive black-box constraints. For the second challenge, we propose a Pareto front-diverse Batch Bayesian Optimization algorithm that selects diverse batches of candidate solutions for parallel evaluation by improving the exploration of the Pareto front. For the third challenge, we introduce non-myopic Bayesian optimization algorithms that plan several steps ahead, balancing immediate gains with long-term benefits. These algorithms are particularly valuable in scenarios with strict resource limitations, where strategic planning can significantly enhance overall efficiency. To demonstrate the practical utility of our proposed algorithms, we apply them to multiple challenging real-world problems, including optimizing 3D printer parameters to produce accurate and precise organ models for presurgical training, discovering high-performing nanoporous materials for hydrogen storage, and designing efficient analog circuits for sustainable computing. Through extensive experiments on synthetic benchmarks and real-world applications, we show that our algorithms substantially improve resource-efficiency to uncover high-quality solutions compared to existing methods. By effectively incorporating user preferences, handling constraints, utilizing batch evaluations, and planning under resource constraints, this dissertation advances the state-of-the-art in multi-objective BO.

Use-Inspired Bayesian Optimization for Engineering Design: From Materials and 3D Printing to Circuits

Files and links (1)

Abstract

Metrics

Details