Skip to the content.

Détails

Organisateurs

Aurélien Alfonsi, Pierre-Cyril Aubin, Alexandre Ern

Programme

Horaire Intervenant(e) / Session
8h45 - 9h15 Accueil café
9h15 - 9h35 Présentation du CERMICS et des invités institutionnels (Gabriel Stoltz)
9h35 - 9h55 Mots d’introduction de Christophe Besse (CNRS/INSMI), d’Anthony Briant (ENPC), de Jamal Atif (IP Paris) et de Pietro Congedo (Inria)

Conférences plénières (10h - 12h25)

Horaire Intervenant(e) Titre
10h - 10h45 Claudia d’Ambrosio (CNRS, LIX) Mixed-integer nonlinear programming
10h50 - 11h35 Thomas Deschatre (EDF) Rainfall is rough
11h40 - 12h25 Greg Pavliotis (Imperial College) Optimizing Langevin dynamics: from convergence to equilibrium to transport coefficients

Déjeuner de 12h30 à 13h45 à la cantine de l’ENPC dans un espace dédié

Exposés du CERMICS (13h45 - 17h15)

Horaire Intervenant(e) Titre
13h45 - 14h15 Laura Grazioli (recrutement CNRS) Differential geometry for the calculation of electronic excited states

Session Pôle Optim

Horaire Intervenant(e) Titre
14h15 - 14h30 Axel Parmentier (Permanent) Decision-focused learning: from industrial applications to rich mathematics
14h30 - 14h45 Héloïse Gachet (Doctorante) Balanced Assignments of Periodic Tasks
14h45 - 15h00 François Pacaud (Ancien doctorant) Logical constraints in nonlinear programming: the case for complementarity constraints

Session Pôle Proba

Horaire Intervenant(e) Titre
15h10 - 15h25 Andrew McRae (Permanent) Random optimization geometry in statistical estimation
15h25 - 15h40 Paul Invernizzi (Doctorant) A study of the reflected kinetic Langevin process
15h40 - 15h55 Clément Hardy (Ancien doctorant) Drug discovery as a multi-fidelity, multi-objective Bayesian optimization problem

Pause café de 15h55 à 16h10

Session Pôle MAS

Horaire Intervenant(e) Titre
16h10 - 16h25 Amaury Hayat (Permanent) The place of AI in the practice of mathematics
16h25 - 16h40 Annamaria Massimini (Post-doctorante) Cross-diffusion in multi-component mixtures: From modeling to numerical approximation
16h40 - 16h55 Jérôme Bonelle (Ancien doctorant) CDO schemes: an EDF–CERMICS collaboration, from mathematics to industrial applications
16h55 - 17h10 Geneviève Robin (Ancienne post-doctorante) Generalizable Machine Learning for Digital Pathology

Clôture

Horaire  
17h15 - 17h30 Mots de conclusion par Yves Bréchet

Résumés

Claudia d’Ambrosio

Mixed-integer nonlinear programming

Decision-making problems are ubiquitous — in energy systems, logistics, engineering design, and finance — and they often combine a large number of discrete choices with nonlinear constraints and objectives. Linear models and heuristics, though widely used, either oversimplify such problems or provide no guarantee of optimality, and general-purpose solvers often fail to scale. Mixed-integer nonlinear programming (MINLP) offers the right mathematical framework, but its full potential remains locked behind hard combinatorial structure. In this talk, I will present a tailored exact method for a class of MINLP problems, provably converging to a global optimum. I will conclude with perspectives and ongoing work on the topic.

↑ Retour au programme

Thomas Deschatre

Rainfall is rough

Résumé disponible sur arXiv : arXiv:2607.27099

↑ Retour au programme

Greg Pavliotis

Optimizing Langevin dynamics: from convergence to equilibrium to transport coefficients

In this talk I will give an overview of some of the work that Tony Lelièvre, Gabriel Stoltz and I have done over the last 15 years, together with other colleagues and former and current PhD students, in particular Urbain Vaes. I will first show how non-reversible perturbations accelerate convergence to equilibrium and reduce the asymptotic variance of ergodic averages. I will then discuss how the diffusion coefficient of reversible dynamics can be optimized with respect to the spectral gap, subject to an appropriate constraint. Finally, I will discuss briefly the calculation of transport coefficients, focusing on a kinetic system of weakly interacting diffusions in a periodic potential. We will study its joint mean-field/diffusive limit and compute the diffusion coefficient of the limiting Brownian motion. The effect of phase transitions of the mean-field dynamics on the diffusion coefficient will be elucidated.

↑ Retour au programme

Laura Grazioli

Differential geometry for the calculation of electronic excited states

One of the most important questions in quantum chemistry is about how to calculate excitation energies of molecules, that is, the energies necessary to promote the system from its ground state to an energetically higher-lying state, as these are necessary to interpret and simulate molecular spectra.

In the many-body Schrödinger theory, electronic excitation energies admit two mathematically equivalent characterisations. On the one hand, excited states arise as higher-energy solutions of the many-body Schrödinger equation and therefore as critical points of the corresponding energy functional; under suitable regularity assumptions, they are saddle points whose number of negative eigenvalues of the Hessian (Morse index) equals the excitation levels. On the other hand, excitation energies appear as the normal modes of the linearised Hamiltonian dynamics around the ground state through linear response theory.

In approximate theories, the two routes to excitation energies do not coincide, making their comparison interesting both from the mathematical and chemical point of view. In this work, we make a step forward by providing a unified geometrical framework, through the investigation of the inherent manifold structures. For the critical point approach, we develop saddle-point search algorithms on the manifold of admissible wave functions. These algorithms target critical points of prescribed Morse index while respecting the intrinsic geometry of the underlying variational manifold, an essential ingredient for the stable computation of inherently unstable saddle points. In parallel, we provide a systematic geometrical derivation of linear response theory for variational wave-function models by exploiting the Kähler structure of the corresponding manifolds. Our derivations reduce to well-known linear response equations in quantum chemistry, which may necessitate ad hoc technicalities for different levels of theory.

We apply the general framework to some of the variational quantum chemistry methods to be able to compare the results for excitation energies and gain insight into the differences between the investigated approaches.

↑ Retour au programme

Axel Parmentier

Decision-focused learning: from industrial applications to rich mathematics

Many industrial decision problems combine combinatorial choices with uncertain, data-rich environments: routing vehicles, scheduling aircraft, designing supply chains, operating power grids. Combinatorial optimization augmented machine learning (COAML) tackles them with policies that chain a statistical model, typically a neural network, with a combinatorial optimization algorithm. The network predicts the parameters of an optimization problem, and the algorithm turns them into a feasible decision. Decision-focused learning (DFL) trains this architecture end to end on the quality of the decisions it produces, rather than on the accuracy of the predicted parameters. This requires learning through an algorithm whose output is a piecewise constant function of its inputs. This short introduction has two goals. First, we give a panorama of applications, from dynamic problems to contextual stochastic optimization. Second, we show that the field leads to rich mathematics at the crossroads of several domains.

↑ Retour au programme

Héloïse Gachet

Balanced Assignments of Periodic Tasks

We address the problem of assigning periodic tasks to workers in a balanced way, i.e., so that each worker performs every task with the same frequency over the long term. Given a list of tasks to be repeated weekly at fixed times and a number of indistinguishable workers, we establish a necessary and sufficient condition for the existence of a balanced assignment when the sole constraint is that no worker performs two tasks simultaneously. Additional periodicity and complexity results are also established for this basic version. In the extended version, additional constraints can be introduced, such as a limit on the total number of working hours per week. For this extended version, it is demonstrated that whenever a balanced assignment exists, a periodic balanced assignment exists as well.

↑ Retour au programme

François Pacaud

Logical constraints in nonlinear programming: the case for complementarity constraints

Many engineering problems, such as security constrained optimal power flow, contact-rich robotics and market clearing, formulate as nonlinear programs with logical constraints. Unfortunately, the traditional reformulation with binary and integer variables yields mixed integer nonlinear programs (MINLP) that are often intractable for modern optimization solvers. Instead, we formulate the logical conditions with complementarity constraints that keep the combinatorial structure explicit while remaining continuous. We solve the resulting problem with a specialized interior-point method, using the celebrated Scholtes relaxation. We introduce in our new solver CCOpt several regularization methods to address the structural ill-conditioning caused by the combinatorial structure. Our numerical results prove that this new solver is effective at solving recent benchmark problems in robotics, energy markets, and power systems.

↑ Retour au programme

Amaury Hayat

The place of AI in the practice of mathematics

AI capabilities in mathematics have seen impressive progress over the last few years. This talk will review how we got there, looking beyond the recent AI-generated results in mathematics, and what the next steps could be.

↑ Retour au programme

Annamaria Massimini

Cross-diffusion in multi-component mixtures: From modeling to numerical approximation

In multi-component mixtures, diffusion dynamics go beyond the classical Fickian picture, in which each component diffuses down its own concentration gradient.

In this talk, I will first introduce the phenomenon of cross-diffusion, explain why it is interesting and important to study, and discuss the analytical and numerical challenges arising from the nonlinear systems of PDEs describing such processes.

I will then focus on the construction of two structure-preserving finite volume schemes for discretizing two different cross-diffusion systems. The first is a two-point flux approximation (TPFA) scheme for a generalized Poisson–Nernst–Planck system on orthogonal meshes; the second is a hybrid finite volume (HFV) scheme for an anisotropic cross-diffusion system on general meshes.

I will conclude by discussing some open questions and possible directions for future research.

↑ Retour au programme

Jérôme Bonelle

CDO schemes: an EDF–CERMICS collaboration, from mathematics to industrial applications

Between 2011 and 2020, three CIFRE PhD projects, supervised by Alexandre Ern and supported by EDF, led to the design and analysis of CDO (Compatible Discrete Operators) schemes and their integration into code_saturne, EDF’s open-source CFD software. This collaboration between CERMICS and EDF gave rise to a new generation of spatial discretization methods designed to preserve the fundamental structures of the governing equations at the discrete level. These developments are now used in several areas of industrial relevance, including subsurface flows, particularly in the context of the Cigéo deep geological repository for radioactive waste, solidification processes, and turbulent flows in the primary loops of nuclear power plants.

↑ Retour au programme

Geneviève Robin

Generalizable Machine Learning for Digital Pathology

In computational pathology, a common task is to analyze digitally scanned tissue samples to predict endpoints that enhance disease understanding and improve patient care. This often involves predicting clinical outcomes such as overall survival (OS) and progression-free survival (PFS), as well as biomarkers like microsatellite instability (MSI/MSS) status. A key requirement for deploying these models in real-world settings is their robustness to variations in preparation protocols (e.g. different stainings) and digitization (e.g., different scanners), ensuring reliability across different clinical centers. State-of-the-art computational pathology pipelines often rely on two components: a large, pre-trained representation model of image patches, also called foundation model, and a predictive model which associates the embedded patches to predicted outcome at the image level. In this work, we present how the robustness and generalizability of both components can be enhanced using model distillation and score calibration.

↑ Retour au programme