Ir. Kurt Sys - Software Architect, QA engineer, Tech Lead Vintecc

Exploiting Uncertainty Outside the Model: Industrial ML as a Hyperliminal Systems Problem

Abstract: In academic data research, exploiting uncertainty is mostly an inside-the-model concern: calibrated posteriors, active learning, optimal experimental design. This talk is about where to embed the model: exploiting uncertainty outside the model, in the environment the model gets deployed into. Industrial deployments are hyperliminal systems: environments where the external forces (stressors) cannot be fully enumerated in advance, where requirements drift faster than they can be specified, and where a perfectly calibrated model still fails because the world around it has moved and there is always "that one exception we didn't think of". In that regime, uncertainty is not a property to minimize; it is the signal that drives every interesting design decision in the system.
The real engineering work, then, is designing the software around the model, the interfaces, the observability and the organizational hooks. The system should encapsulate the model and shield it from a complex and often uncontrollable environment. That means thinking carefully about how guidance and explanation reach the user, accepting that you can never fully trust an end user to interpret a result correctly, and treating every analysis as a default lie until the surrounding system proves it true. Those default lies are not a flaw to be designed out; they are the load-bearing assumption that keeps the design honest, and the driving force behind every robust and resilient industrial system we build.

Biography: Kurt is a recovering microbiological systems engineer. At Vintecc he is the tech lead and architect on the Data & AI Ops team: broadly, the plumbing around the model that nobody asks for and everybody needs. Previously: kept astronauts alive in long-term space travel using microbiological life support systems, where the microbiology is inherently complex but the environment controlled and bounded. Currently: keeps analytical models alive on factory floors, where the model is simple and the environment is neither. Reports the second is harder.

Dr. Pieter Muyshondt - Senior Engineer Simulation and Analysis Cochlear

Understanding variability of middle-ear motion through simulation

Abstract: In the development of hearing implants, predicting and understanding how a device will perform across the patient population is crucial. The conduction of sound to the hearing organ is governed by highly complex 3D vibrations of the eardrum and ossicles, showing a strong variability between individuals. Understanding which components in the ear anatomy are responsible for the vibration modes and the variation between subjects is challenging, given the limited availability of measurement data of both motion and structure. This presentation will focus on how finite element analysis of the middle ear is used together with parameter sensitivity analyses to replicate measurements of 3D ossicular vibrations collected with 3D laser vibrometry, and to improve understanding in the sources and nature of the observed variation.

Biography: Pieter Muyshondt is a senior simulation engineer at Cochlear Technology Centre Belgium and former postdoc researcher at the Laboratory of Biomedical Physics at University of Antwerp. He has a PhD in Physics on numerical simulation and characterization of middle-ear mechanics. His main interests are finite-element analysis, vibroacoustic testing and data analysis in hearing research and technology, focusing on the development of future hearing implants.

Prof. dr. Bram Vervisch - UGent, ORBITS

Beyond predictive maintenance:Prioritizing the 'why' over the ‘when' in electric drivetrains

Abstract: While predictive maintenance has emerged as a highly sought-after objective across various industrial sectors—offering a compelling theoretical alternative to reactive and periodic maintenance approaches—its primary focus is often misdirected. Currently, both academia and industry are investing heavily in advanced measurement techniques and modeling to predict exactly when an industrial asset will fail. However, this presentation argues that this objective is fundamentally insufficient. Even if forecasting the precise time of failure were perfectly accurate, understanding why a component fails is of far greater strategic importance.

A vast majority of faults in electrical drive systems are entirely preventable if the underlying causes are addressed. Therefore, this presentation shifts the focus toward the fundamental failure behavior of electrical machines. We will demonstrate how this behavior can be systematically mapped and understood using a Reliability Impact Analysis (RIA). Furthermore, we argue for a paradigm shift in condition monitoring techniques: methodologies must prioritize identifying the root cause (the "why") over merely predicting the failure timeline (the "when"). Finally, we will introduce the Machine Health Scan developed by ORBITS as a practical, industry-ready implementation designed to achieve this proactive, root-cause-driven objective.

Biography: Bram Vervisch is a distinguished expert in mechanical vibrations, complex machinery diagnostics, and predictive maintenance. As the co-founder and managing director of ORBITS, he’s part of a specialized team that provides advanced diagnostic solutions to international companies in demanding sectors such as petrochemicals, pharmaceuticals, and offshore. His current focus lies in guiding the company's strategic vision and driving technical innovation. Alongside his role at ORBITS, he shares his extensive knowledge as a guest professor at Ghent University (UGent), where he teaches Machine Optimization and Machinery Diagnostics. Holding a PhD in rotordynamics from UGent and an ISO 18436 CAT III certification, Bram leverages his strong academic background in classical vibration and Electrical Signature Analysis (ESA) to oversee the translation of complex theoretical data into sustainable, high-level industrial solutions.

Drs. Gabriel Diaz-Aylwin - University of Lancaster, United Kingdom Atomic Energy Authority

Bayesian Optimisation for fusion reactor design

Abstract: I will discuss approaches to adapting the traditional Bayesian optimisation loop to the structured but challenging physical environment of tokamak fusion reactors.

In particular, two challenges arise:

• First, we have to handle constraint functions on the state, i.e. the solution of a PDE, as opposed to the parameters defining its boundary conditions or residual operator.
• Second, we have a range of objective functions each with varying fidelity and corresponding run time varying from seconds to weeks.

I will address both challenges by constructing geometry-trust regions and integrating multi-fidelity modelling into the optimisation loop.

Biography: Having previously studied mathematics and theoretical physics at the University of Oxford and worked on practical machine learning problems in aerospace, I am presently pursing a PhD in applied mathematics at the University of Lancaster. Here, I work on computational approaches to engineering design optimisation. Specifically, my project centres around divertor optimisation in fusion reactors and is sponsored by the UK Atomic Energy Authority.

Dr. Dmitry Bagaev - TU/e, former CTO Lazy Dynamics

Keep Calm and Trust AI

Abstract: Large language models are powerful, but when asked for facts or numerical answers, they can hallucinate with surprising confidence. In this talk, we take a different approach: instead of asking an LLM to guess, we let it orchestrate real probabilistic inference. Using RxInfer and an MCP server, we connect a language interface to a Bayesian linear regression model running on an actual dataset. The LLM translates user intent into structured computation, the regression model performs principled inference, and the result is grounded in data—not generated from patterns alone. You’ll see how combining probabilistic programming with tool calling creates AI systems that are transparent, verifiable, and dramatically more reliable. Sometimes, the best way to trust the AI is to make sure it does the math.

Biography: Senior software engineer and PhD scientist with a strong mathematical foundation and expertise in software development, machine learning and data science. Brings a unique blend of academic rigor and hands-on industry experience, with a proven track record of leading technical teams, architecting complex systems, and translating cutting-edge research into practical applications.

Dr. Peter Verleijsdonck - TU/e, BiasLab

Improving Energy Management Systems for Congestion Challenges through Deep Reinforcement Learning and Bayesian Inference

Abstract: Grid congestion is accelerating the deployment of battery energy storage systems (BESS), which provide peak shaving during high-demand hours and grid-balancing services at other times. Balance Responsible Parties (BRPs) manage portfolios of BESS using smart Energy Management Systems to deliver these services efficiently. In the imbalance market, BRPs trade deviations from scheduled generation or consumption to maintain system balance. Real-time price signals from the grid operator create financial incentives that can shorten BESS investment payback periods. In collaboration with Zympler, a BRP in the Netherlands, we develop a dual AI approach to optimize energy trading for BESS. First, we use RxInfer to learn a generative model of market dynamics via Bayesian inference. Next, we apply DynaPlex to optimize battery operations through deep reinforcement learning based on the learned model. Using the complex price signals of the Dutch imbalance market as a case study, we demonstrate how this dual AI approach can accelerate return on investment.

Biography: Peter is a researcher in the Department of Electrical Engineering at Eindhoven University of Technology (TU/e), where he works on probabilistic AI for flexible energy management. His current research develops Bayesian AI agents that optimize the use of flexible energy resources and strengthen the reliability of the Dutch energy networks. He earned both his MSc and PhD in Applied Mathematics from TU/e. His work centers on building models of reality grounded in data and physical interpretation to address everyday problems in logistics, maintenance, and computer science, with a particular focus on stochastic models and simulations for network, queueing, and decision-making problems. Alongside this, he holds deep interests in computer science and machine learning algorithms.