Machine learning and physics strengthen each other, scientists say

Machine learning is moving from the margins of particle physics to its very center, and two of Germany’s leading theoretical physicists argue that this transformation will succeed only if the field defends the methodological pillars that made it rigorous in the first place. In an editorial published in The European Physical Journal C, Michael Krämer of RWTH Aachen University and Tilman Plehn of Heidelberg University lay out a detailed assessment of how artificial intelligence is reshaping fundamental physics—enhancing classic analyses, enabling entirely new ones, and, in turn, posing hard technical questions that machine learning researchers have largely ignored. Their central message is one of cautious enthusiasm: AI is an opportunity physics cannot afford to reject, but it must never be allowed to redefine what counts as a discovery.

The authors begin from a deceptively simple premise. The goal of fundamental physics has never been to describe data; it has been to identify abstract mathematical structures that explain and relate observations across wildly different systems and energy scales, from particle collisions to cosmology. In particle physics this program is explicitly reductionist: complex phenomena are assumed to become simpler when expressed in terms of elementary degrees of freedom and their interactions. Historically, that philosophy steered experiments toward well-controlled, simplified setups—electron-positron colliders ahead of proton colliders, dedicated dark matter experiments, and other systems where reliable theoretical predictions were tractable. But modern datasets have outgrown that comfort zone. The precision program at the Large Hadron Collider and the vast radio surveys planned for the Square Kilometre Array both produce data far too complex for traditional summary statistics such as low-dimensional histograms or power spectra. The promise of machine learning, in this framing, is to provide near-optimal data representations and inference strategies that combine huge datasets with first-principles theory simulations in a statistically controlled way.

Krämer and Plehn are careful about what machine learning is and is not. It does not, for now, change the scientific objectives of the field. Particle physics, because of its quantum nature, relies on statistical inference rather than deterministic validation, and that framework is non-negotiable: an experimental observation must be statistically validated and, in principle, connected to a broader theoretical description. A discovery made in one experiment or at one energy scale must be testable elsewhere. Machine learning, the authors insist, extends the set of tools connecting theoretical models to data—it does not replace the theoretical and statistical scaffolding within which physics results are interpreted.

The technical heart of the editorial is a survey of how machine learning is already transforming every stage of the classic particle physics analysis chain. Triggering systems at the LHC are becoming more efficient and flexible through ultra-fast neural networks evaluated in real time. Data acquisition and the fusion of detector information increasingly rely on learned latent representations. Object identification and classification—the original success story of deep networks in the field, from jet images to particle-cloud taggers like ParticleNet—have become standard examples of machine learning outperforming established methods, with detector calibration a natural next target. On the theory side, neural networks simplify analytic expressions, solve differential equations, and accelerate event generators. Generative machine learning, the authors write, transforms theory simulations not at the cost of their first-principles nature, but by enabling predictions at levels of complexity where numerics are severely limited.

Beyond these enhancements, machine learning is enabling analyses that were previously impractical. Anomaly searches are the flagship example. Their weakly supervised variants generalize the classic bump hunt, in which a histogrammed observable is fitted everywhere except a sliding signal window; a neural network now learns the background model for potential signal regions. The authors stress that such methods are rarely entirely new—they build on existing concepts that were not viable without machine learning, and their real difficulty lies in statistics, from controlling the look-elsewhere effect to understanding the systematics of network training. Autoencoder-based anomaly scores, well established in the machine learning community, will monitor detector performance and improve triggers, but will only become fully useful for physics once embedded in a consistent statistical framework. A second direction is analysis reinterpretation through unfolding and deconvolution, where machine learning’s advantage in high dimensions could turn reinterpretation into a standard method for making experimental data publicly available.

The most conceptually far-reaching development is representation learning: the transition from predefined feature spaces to learned latent representations. Because particle physicists know the underlying symmetry structure of their data, optimal representations benefit from explicitly equivariant architectures—Lorentz-equivariant transformers, for example—as well as implicitly learned symmetries. The authors note that the strict division between explicit and implicit symmetry treatment dissolves in practice, since the symmetries relevant at experiments are typically broken, leaving open the question of whether to learn the breaking of a known symmetry or the broken symmetry from scratch. The deep open problem, they argue, is connecting learned latent spaces to the theory-driven representations of quantum field theory, where the simulation chain is a sequence of factorized conditional probabilities rooted in the Lagrangian. If a machine learning method spotted an unknown pattern in the latent space of LHC data, that alone would not be a discovery; a genuine discovery demands a statistically validated background model and a generalizing theory prediction—perhaps, they suggest, extremely weakly interacting particles appearing in semi-visible jets, followed by quantum field theory work showing the signal describes dark matter production in the early universe.

The editorial then turns the tables, asking how physics challenges machine learning. Particle physics treats classification as learning class probabilities, demanding calibrated uncertainty estimates that are uncommon in standard benchmarks. It asks generative networks trained on a million events how many samples can be drawn before training statistics run out—a question natural to physics, rare in machine learning. It demands control of biases from architectures, hyperparameters, and learned representations, far beyond typical practice, as illustrated by neural-network parton distribution functions of the proton. Information geometry, with geodesic distances and curvature applied to latent spaces, offers a route to quantitative explainability that draws directly on physics concepts, while complex-systems ideas and statistical field theory are being applied to understand how neural networks learn at all.

The authors are blunt about the dangers. First, machine learning must not be used outside a proper statistical framework; the more powerful the method, the more critical its statistical control. Second, machine-defined “theory” models that describe data well must not replace first-principles theory, because only a framework like quantum field theory can connect LHC measurements to nuclear physics, flavor physics, or cosmology and generalize beyond the training data. Third, an inflation of AI-accelerated analyses with limited novelty threatens to blur what counts as a scientific contribution; running a powerful workflow on another dataset will not suffice. Yet their response is not conservative retreat. AI will accelerate analyses, improve simulations, enable flexible new methods, and make complex workflows transparent—including through agentic systems that orchestrate simulation, inference, and interpretation using domain-specific tools. Tool-augmented agents navigating event generators and likelihood evaluations fit within the scientific method; whether general-purpose language models can generate meaningful physics knowledge on their own remains an open question.

The editorial closes with a reminder that most fundamental physics happens at universities, where the societal contribution includes training young scientists who will carry analytical and data skills into industry. Teaching scientifically sound AI methods—how to apply them, understand their limits, and control their uncertainties—must become part of the field’s AI strategy, alongside attention to the energy footprint of generative tools. The AI transformation, the authors conclude, is happening faster than any historical precedent. Fundamental physics must embrace it while holding its statistical and theoretical standards firm—and if it succeeds, AI will not only change how physics is done, but help define what physics asks next.