Active Learning & Optimization

Research Area

Active Learning and
Optimization under Uncertainty

We develop active learning and advanced optimization methods that scale to system-level digital twins with inherent noise for data-efficient and interpretable decision-making across large-scale engineering systems.

Overview

Optimizing real engineering systems requires integrating knowledge from multiple sources — reduced-order models, high-fidelity simulations, and physical experiments — each carrying its own form of uncertainty. Aleatoric uncertainty (inherent noise in measurements and processes) and epistemic uncertainty (limited data and model approximations) both shape what solutions are feasible and how reliable they are in practice.

Our research develops optimization frameworks that embrace this heterogeneity: combining data from models and experiments of varying cost and fidelity, explicitly accounting for noise and uncertainty in constraints and objectives, and scaling to problems that span multiple engineering disciplines. The goal is to find designs that perform well not just in simulation, but robustly in the real world.

These methods form the optimization backbone of the broader DISO Lab research portfolio, supporting digital twin-enabled co-design, real-time decision-making, and lifecycle optimization of autonomous engineering systems.

Key Research Themes

Bayesian Optimization

We use uncertainty-aware machine learning models and acquisition functions to efficiently navigate high-dimensional design spaces under aleatoric uncertainty (inherent measurement and process noise), finding robust optimal solutions with far fewer evaluations than classical methods.

Multi-Objective Optimization

We develop parametric multi-objective optimization methods that map design parameters to entire Pareto fronts, enabling engineers to understand performance tradeoffs. We additionally design surrogate-assisted constraint-handling techniques that predict feasibility probability, ensuring constraint satisfaction even when evaluations are noisy or expensive.

Co-Design & Multidisciplinary Design Optimization

We formulate and solve optimization problems that span more than one engineering domain simultaneously — such as structural, control, thermal, and manufacturing disciplines — enabling holistic system designs that no single-discipline approach can achieve.

Multi-Fidelity Active Learning

We develop active learning strategies that intelligently combine reduced-order models, high-fidelity simulations, and physical experiments — each at a different cost and accuracy — to allocate computational and experimental budgets where they yield the greatest improvement in the surrogate model and optimization outcome.

Selected work

Related Publications

Adaptive uncertainty-aware material-process co-design for in-space manufacturing

Optimization 2026

Adaptive Uncertainty Aware Material-Process Co-Design for Sustainable In-Space Manufacturing of Recycled Composites

Ying-Kuan Tsai†, Christian A. Fernandez†, Ryan Daniel Brahan, Anupam Ajit Deshpande, Caleb Isaiah Smith, Yayue Pan, Wei Chen

(†Indicates Equal Contribution)

ASME IDETC-CIE Conference · Houston, Texas · August 2026 (accepted)

Optimization 2025

An Optimization-Centric Review on Integrating Artificial Intelligence and Digital Twin Technologies in Manufacturing

Vispi Karkaria, Ying-Kuan Tsai, Yi-Ping Chen, Wei Chen

Engineering Optimization · January 2025

Paper

Optimization 2024

Surrogate‑Assisted Constraint‑Handling Technique for Parametric Multi‑Objective Optimization

Ying-Kuan Tsai, Richard Malak

Structural and Multidisciplinary Optimization · September 2024

Paper Code

Approximate explicit model predictive controller design

Optimization 2022

Design of approximate explicit model predictive controller using parametric optimization

Ying-Kuan Tsai, Richard J Malak Jr

Journal of Mechanical Design · December 2022

Surrogate modeling for sonic boom minimization in morphing aircraft