Professor Daniel N. Wilke

The Art and Science of Mismatch: Aligning Data, Models, Latent Spaces, Fidelity and Engineering Objectives

University of the Witwatersrand

Professor Daniel N. Wilke is a South African researcher and educator specialising in physics-based, hybrid and data-driven generative modelling and optimisation.

His work includes the development of gradient-only optimisation methods for discontinuous loss functions, multi-fidelity surrogate modelling strategies, data pre-transformation strategies and improvements in latent variable models to enhance interpretability and causal insights. Through his research, teaching, and consulting, he consistently applies theoretical insights to address practical challenges in engineering and education.

He served as President of the South African Association for Theoretical and Applied Mechanics and as National Chairman of the International Union for Theoretical and Applied Mechanics.

He has authored two books—Springer’s Practical Mathematical Optimization and The Lost Art of Learning—the former having received the Vice-Chancellor’s University of Pretoria Book Award. He is currently the Head of the School of Mechanical, Industrial and Aeronautical Engineering at the University of the Witwatersrand of Johannesburg.

Abstract

The Art and Science of Mismatch: Aligning Data, Models, Latent Spaces, Fidelity and Engineering Objectives

The art and science of mismatch examine how mismatches arise when data, models, latent spaces, and engineering objectives fail to align. In addition, modern data science often trains on data collected at different levels of fidelity, each with their own form of epistemic uncertainty, in addition to aleatoric uncertainty. These variations create extra layers of complexity when selecting and training models that must compress, interpret, or generate information.

In addition, practitioners can often easily express their goals verbally but are required to express them mathematically resulting in proxies to define metrics, latent spaces, optimisation objectives and constraints. These proxies guide model fitting, support surrogate modelling, and guide the distinction between variance-driven or interpretation-driven latent variable models. Yet these proxies are not the true goals of an engineering or scientific task.

The art and science of mismatch explore foundational ideas for identifying, understanding and resolving these mismatches. In addition, to providing a simple conceptual map for working with imperfect data, mismatched models, and the practical demands of engineering design and decision-making. The aim is to develop an appreciation that through a careful selection and alignment of data, fidelity, models, objectives, and latent structure we can develop more efficient, transparent and purposeful models.

Back to speakers