Hi, I'm Raphaël 👋

We present a first application of sparse-autoencoder-based mechanistic interpretability to particle physics. Studying a neutrino foundation model pretrained on IceCube data and fine-tuned for direction reconstruction, we identify a validated atlas of physical concepts in the model representation, using a strict validation protocol consisting of held-out tests, matched nuisance controls, and replication across independent dictionary trainings. Causal interventions show that the direction head barely draws on this atlas. Motivated by this underused information, we train an uncertainty head on the same event-level representation to predict the model's angular reconstruction error. Unlike the direction head, it depends causally on quality and brightness features from the atlas. At $20\%$ selection efficiency, this interpretable estimator improves the median angular resolution from $20.2^\circ$ to $3.2^\circ$. These results suggest that mechanistic interpretability can reveal learned latent physics encoded within a model's internal representation and help design downstream tasks that exploit it.

In many classification problems, reliable instance-level labels are unavailable. However, it is often possible to construct weakly enriched unlabeled samples: datasets selected by different cuts, sources, populations, or experimental conditions that change latent class proportions without revealing them. Classification without Labels (CWoLa) shows that, in the binary case ($K=2$), a classifier trained to distinguish two impure mixtures with different class proportions can recover an optimal class discriminator without knowing the mixture proportions. We extend this principle to multiclass learning from several unlabeled mixtures ($K>2$), where the learner observes only mixture identity and neither latent class labels nor class-prior matrices. We prove that, for a multiclass mixture model, the Bayes-optimal mixture classifier $g^\star$ maps data points into a $(K-1)$-simplex embedded in mixture-posterior space. Leveraging this geometry, we propose prior-free procedures that train a standard classifier to distinguish mixture identities and then extract latent class structure using either post-hoc simplex fitting or a bottleneck architecture. Experiments on MNIST, CIFAR-10, and Galaxy10 DECaLS show that mixture identity alone can recover latent classes and their fractions in the mixture.

We show that Shapley values can be used to trace how individual parton distributions (PDFs) shape the theory predictions for high-energy observables computed from them. This provides a tool for assessing the impact of data on PDFs when determining them, and the impact of PDF uncertainties when using the PDFs to compute collider observables. The Shapley value is computed by treating the regression of PDFs from data as a cooperative game. The PDFs are the players, and the reward is the likelihood ($\chi^2$) that characterizes the agreement between data and the predictions obtained from a given PDF, with theory and methodology held fixed. We find that the method recovers known expectations about which data constrain which PDFs in a global fit, while placing them on a more quantitative footing. We uncover an unexpected loss of sensitivity of the gluon PDF at intermediate $x$, with potential implications for BSM searches and the gluon fusion Higgs cross section.

Tracing the contribution of individual parton distribution function (PDF) flavors to the χ² of specific datasets is difficult due to the convolutions, basis rotations, and DGLAP evolution involved in the PDF-to-observable pipeline. We propose to address this using Shapley values, assigning each flavor its average marginal contribution to the χ² for a chosen dataset and x-region. The small number of PDF flavors makes exact computation feasible, avoiding the feature-independence approximation of existing frameworks. Applied to NNPDF4.0, the method correctly recovers the known flavor hierarchy and identifies which datasets and kinematic regions drive the fit.

In the current LSST survey, transient detection including events like Type Ia supernovae (SNe Ia)—will be conducted through Difference Image Analysis (DIA). A major challenge in this method is that many detections are actually bogus, arising from noise, artifacts, imperfect image subtraction, cosmic rays, bad pixels, or atmospheric effects. In this talk, we introduce a machine learning-based approach for classifying bogus and transient events using unlabelled datasets. By injecting synthetic transients into the data, we eliminate the need for human labelling. Turning the problem into a weakly supervised machine learning task, we present our adapted method Asymmetric Co-Teaching. We also display how to propagate uncertainties using a mixture of Ensemble and MC-dropout. Finally we show interactive plots of latent space projections that confirms our uncertainties.

The presentation began with an overview of interpretability methods, followed by three applications. The first application focused on uncertainty quantification and latent space exploration for real/bogus classification in the context of supernova detection at the Vera Rubin Observatory. The second application presented initial results using an adapted explainable AI method, Shapley values, in the study of parton distribution functions (PDFs), with the aim of mapping the impact of individual PDFs on the fit, specific datasets, and particular x-regions. Finally, the talk concluded with a future project exploring the use of mechanistic interpretability for physics foundation models, such as those used for neutrino detection, with the goal of highlighting the underlying physical reasoning learned by the model.