Engineered Inference Systems: From Empirical Performance to Physical Admissibility, Diagnosis, and Re-Certification

Speaker:  Dr. Lorenzo Jr. Sabug

13 Aprile 2026 | 10:00
DEIB, Sala Conferenze "E. Gatti" (Ed. 20)

Contatti:  Prof. Fredy Orlando Ruiz Palacios

Sommario

On April 13^th, 2026, at 10:00 am Dr. Lorenzo Jr. Sabug, Research associate, Department of Electrical and Electronic Engineering, Imperial College London, will give a seminar on "Engineered Inference Systems: From Empirical Performance to Physical Admissibility, Diagnosis, and Re-Certification" in DEIB Conference Room "Emilio Gatti" (Building 20).

Inference systems, from classical statistical estimators to modern machine learning, are increasingly deployed in high-stakes physical systems.
However, their qualification still relies heavily on modelling assumptions, empirical performance, and post hoc testing. When those assumptions are violated, such systems may continue operating while producing internally consistent but physically inadmissible outputs. Without explicit checks on physical consistency, this can drive unsafe decisions, as seen in aerospace incidents where such failures are mission-critical. This exposes a structural gap: existing inference methods are typically designed to estimate under assumptions, but not to make validity explicit, enforce physical admissibility, or respond systematically when consistency is lost. This talk introduces a perspective of engineered inference systems: architectures developed as engineered artefacts with explicit assumptions, admissibility-aware operation, diagnosable failure modes, and principled re-certification procedures. Rather than advocating a single model class, the emphasis is on a lifecycle view in which inference systems are designed not only to perform, but to remain physically valid and auditable in deployment. To ground these ideas, I present Direct Constraints-Based Regression (DCBR) as a proof-of-existence framework that replaces training loops with constrained optimisation, enabling inference over admissible sets and explicit failure when consistency cannot be maintained. I then outline the broader research direction this motivates: a new engineering assurance standard for inference systems, spanning design, operation, diagnosis, repair, and re-certification appropriate for mission-critical applications.

Biografia

Lorenzo Sabug, Jr. is research associate at Imperial College, where he is co-investigator and intellectual lead for the UK Research & Innovation EPSRC grant "Concurrent Learning and Control of Uncertain Large-Scale Phenomena". He currently investigates regression methods for physics-informed machine learning and dual control, and their applications on propagation-based phenomena. He was previously postdoctoral research fellow at Politecnico di Milano, where he also earned a Ph.D. (cum laude) in systems and control, and has worked on mathematical and data-driven optimisation methods, and their industrial and aerospace applications. His research interests are in set-theoretic methods, and inference systems for mission-critical autonomy.