We propose physics-informed neural networks (PINNs) as a numerical solver for relativistic spin hydrodynamics and demonstrate that the total angular momentum, i.e., the sum of orbital and spin angular momentum, is accurately conserved throughout the fluid evolution by imposing the conservation law directly in the loss function as a training target. This enables controlled numerical studies of the mutual conversion between spin and orbital angular momentum, a central feature of relativistic spin hydrodynamics driven by the rotational viscous effect. We present two physical scenarios for a rotating fluid confined in a cylindrical container: one case in which initial orbital angular momentum is converted into spin angular momentum in analogy with the Barnett effect, and the opposite case in which initial spin angular momentum is converted into orbital angular momentum in analogy with the Einstein–de Haas effect. Assuming translational invariance along the axial direction and a rotational symmetry, we reduce the dynamics to an effective ($1+1$)-dimensional problem in time and radial coordinates and investigate these conversion processes governed by the rotational viscous effect by analyzing the spacetime profiles of thermal vorticity and spin potential. Our PINN-based framework provides the first numerical evidence for spin-orbit angular momentum conversion with fully nonlinear computational relativistic spin hydrodynamics.

Large language models and agentic systems have recently demonstrated potential for automating scientific workflows, including atomistic simulations. However, their deployment in high-performance computing (HPC) environments remains limited by the lack of mechanisms ensuring correctness, reproducibility, and safe interaction with computational resources. Generated workflows suffer from inconsistencies, incorrect application programming interface usage, or invalid physical configurations—leading to failed or unreliable simulations. In this work, we introduce lara-hpc (Lara Agentic Research Assistant for HPC), a validation-driven agentic framework to enable reliable workflow generation for atomistic modeling on HPC systems. Our approach is based on three key components: (1) a controlled execution layer that mediates all interactions with HPC resources; (2) simulation-native validation through dry-run capabilities, enabling execution-level verification without incurring resource cost; and (3) a multiphase agentic pipeline combining retrieval-augmented generation and iterative refinement. We demonstrate the effectiveness of this approach by performing an end-to-end atomistic simulation workflow on HPC by applying lara-hpc to density functional theory simulations. The results show that validation-driven generation significantly improves robustness and enables iterative correction of both syntactic and physical inconsistencies. More broadly, this work advocates for a shift from generation-first to validation-first paradigms in artificial-intelligence-assisted scientific computing. We argue that the future task of the computational physics community is to develop domain-specific agentic systems based on structured tooling to realize an HPC enabled copiloted research ecosystem.

Autoregressive generative models—including Transformers, recurrent neural networks, classical Kalman filters, state space models, and Mamba—all generate sequences by sampling each output from a deterministic summary of the past, producing genuinely non-Markovian observed processes. We develop a general theoretical framework based on stochastic thermodynamics for this class of architectures and introduce the entropy production, which can be efficiently estimated from sampled trajectories without exponential overhead, despite the non-Markovian nature of the observed dynamics. As a proof-of-concept experiment with a large language model (LLM), we evaluate the entropy production for a pretrained Transformer-based model, GPT-2. We find that the token-level entropy production is dominated by a syntactic artifact, while the sentence-level entropy production tends to be larger for causally ordered than for noncausal text sets. This observation is supported by a reevaluation with a substantially larger model, Qwen3-4B-Base. We also demonstrate the framework in the linear Gaussian case, where the model reduces to the Kalman innovation representation and the entropy production admits an analytical expression. We also show that the entropy production decomposes exactly into non-negative per-step contributions in terms of retrospective inference, and each of those terms further splits into information-theoretically meaningful terms: a compression loss and a model mismatch. Our results establish a bridge between stochastic thermodynamics and modern generative models, and provide a starting point for using irreversibility as a quantitative probe of the highly non-Markovian processes generated by models such as LLMs.

The workflow from particle collision to physics analysis passes through a series of reconstruction steps that are traditionally modular and disconnected, with no shared representation linking low-level detector data to high-level analysis tasks. We show that casting event reconstruction as a machine learning problem naturally produces such a shared representation. We repurpose a machine learning model trained for particle-flow reconstruction (MLPF) to perform three distinct analysis tasks: jet flavor identification, jet energy regression, and missing momentum regression. By appending the per-particle latent representations learned during reconstruction as additional input features, we substantially improve over baselines that use conventional particle-level features alone. We further demonstrate that a single linear layer trained using only the latent representations achieves competitive performance against state-of-the-art baseline architectures, and outperforms the baseline for missing momentum regression with approximately 35 times fewer parameters. These results demonstrate that the latent representations learned during reconstruction encode essential physics information needed for downstream analysis, establishing MLPF as a foundation model and offering a concrete step toward an end-to-end pipeline from detector data to physics analysis.

A key question in cell biology is how cell-scale organization emerges from a given set of molecular players and rules of interaction. Given its multiscale nature, addressing this question requires a combination of experimental perturbation, mathematical modeling, and parameter inference. We leverage recent advances in each of these fields, focusing, in particular, on neural-network methods for simulation-based inference, to study how cell-scale patterns of Rho guanosine triphosphatase (GTPase) activity are defined by molecular-scale activator-inhibitor interactions with filamentous actin. Using an existing model of this interaction, we demonstrate that an overexpressive, regularized classification neural network can approximate the likelihood of data arising from a particular parameter set. We show that variations in F-actin assembly dynamics can be inferred directly from experimental data, but only if the network is made less sensitive to model misspecification. We use our approach to interpret perturbation experiments in which increasing Rho GTPase-activating protein (RhoGAP) coexpression increases the frequency and coherence of Rho activity waves in frog eggs. After showing that the known functions of RhoGAP are insufficient to explain experimentally observed dynamics, we use neural methods to suggest an alternative pathway by which RhoGAP could decrease filament nucleation rates to sustain waves. Our work yields specific, experimentally testable predictions and illustrates how a combination of traditional forward simulation models and modern inference tools can aid in unraveling mechanisms of self-organization.

Artificial intelligence (AI) has become an exceptionally powerful tool for analyzing scientific data. In particular, attention-based architectures have demonstrated a remarkable capability to capture complex correlations and to furnish interpretable insights into latent, otherwise inconspicuous patterns. This progress motivates the application of AI techniques to the analysis of strongly correlated electrons, which remain notoriously challenging to study using conventional theoretical approaches. Here, we propose AI workflows for analyzing snapshot datasets from tensor-network simulations of the two-dimensional (2D) Hubbard model over a broad range of temperature and doping. The 2D Hubbard model is an archetypal strongly correlated system, hosting diverse intriguing phenomena including Mott insulators, anomalous metals, and high-$T_c$ superconductivity. Our AI techniques yield fresh perspectives on the intricate quantum correlations underpinning these phenomena and facilitate universal omnimetry for ultracold-atom simulations of the corresponding strongly correlated systems.

Recent variational studies have demonstrated that the strongly correlated ground states of the fractional quantum Hall (FQH) effect can be captured using machine learning approaches starting from no prior knowledge of the underlying physics. We introduce a complementary framework that instead starts from Jain's composite-fermion (CF) wave functions, which accurately describe FQH states as weakly interacting states of CFs at fillings $\nu =n/(2pn+1)$ in an idealized limit. As we move away from this idealized limit to one more in line with experimental reality, we expect CFs to become dressed much like the electrons of a noninteracting system, which are dressed by neutral excitations as the interaction is turned on adiabatically, as in Landau's Fermi-liquid theory. We model this dressing using a Feynman-Cohen-style backflow approach, implemented through symmetry-preserving neural networks, which we refer to as CF-Flow. CF-Flow achieves competitive accuracy with substantially greater computational efficiency and scales to systems of $\gtrsim 26$ electrons. At fillings $\nu =1/3$ and $2/5$, CF-Flow produces ground-state energies as functions of Landau-level mixing strength that are nearly indistinguishable from those obtained using the fixed-phase diffusion Monte Carlo (fp-DMC) method, with low local-energy standard deviations, even though the latter constrains the wave function phase to that of the lowest Landau level, thereby providing insight into why fp-DMC has been successful in giving an accurate quantitative account of several experiments. Finally, the symmetry-preserving architecture of CF-Flow enables access to excited states and computation of the transport gap at $\nu =1/3$ and $\nu =2/5$. Within the FQH sectors studied, the extrapolated gap is well described by an exponential decrease with Landau-level mixing and approaches a finite value at large mixing strength.

Large language model (LLM)-driven agents are emerging as a powerful new paradigm for solving complex problems. Despite the empirical success of these practices, a theoretical framework to understand and unify their macroscopic dynamics remains lacking. This paper proposes a method based on the least action principle to estimate the underlying generative directionality of LLMs embedded within agents. By experimentally measuring the transition probabilities between LLM-generated states, we report evidence consistent with a tendency toward detailed balance in LLM-generated transitions, indicating that LLM generation may not be achieved by generally learning rule sets and strategies, but rather by implicitly learning a class of underlying potential functions that may transcend different LLM architectures and prompt templates. To our knowledge, this provides an effective macroscopic description of LLM generative dynamics that does not depend on specific model details. This work is an attempt to establish a macroscopic dynamics theory of complex AI systems, aiming to elevate the study of AI agents from a collection of engineering practices to a science built on effective measurements that are predictable and quantifiable.

We introduce a variational wave function based on neural-network quantum states (NQS) to study lattice systems whose local Hilbert space contains both spin and fermionic degrees of freedom. Our approach is based on the use of the transformer architecture, which can naturally handle composite local Hilbert spaces through a tokenization procedure closely inspired by techniques from natural language processing. The key methodological innovation is that the neural network predicts a set of fermionic orbitals that depend on the spin configuration in a backflow-inspired manner. We apply the method to the one-dimensional ancilla layer model, consisting of a chain of mobile spin-$1/2$ fermions coupled to a two-leg spin-$1/2$ ladder. For open boundary conditions, we achieve excellent quantitative agreement with density matrix renormalization group results across the full range of parameters considered. We find a phase in which the chain forms an effectively decoupled Luttinger liquid (LL), and an ${\mathrm{LL}}^{*}$ phase with a distinct Fermi wave vector in which the mobile fermions are Kondo screened by one leg of the ladder, while the other leg forms the critical Bethe spin liquid. The ${\mathrm{LL}}^{*}$ is the analog of the ${\mathrm{FL}}^{*}$ phase in two dimensions proposed to describe the cuprate pseudogap. We also find a Luther-Emery phase, where the ${\mathrm{LL}}^{*}$ state becomes unstable toward the formation of a spin gap. The transformer ansatz maintains comparable accuracy for periodic boundary conditions, where tensor-network methods are computationally more demanding. Moreover, we show that the approach extends naturally to fully periodic two-dimensional clusters, a regime beyond the reach of current tensor-network methods. Together, these findings establish transformer-based NQS as an accurate and scalable variational framework for correlated lattice systems with composite local Hilbert spaces and highlight its potential for studying higher-dimensional models where boundary effects and heterogeneous local structures pose significant challenges.

There has been a veritable explosion of methods and software to perform machine-learned regression on datasets of electronic energies and forces to develop high-dimensional machine-learned potential energy surfaces (ML-PESs). A major, but not deeply studied aspect is how well different ML-PESs represent the same dataset on which they were trained, beyond the standard statistical precision measures. Here, this is examined using several “stress tests,” for two widely used machine-learned potential approaches. One is based on permutationally invariant polynomial linear least-square regression and the other is the message-passing neural network PhysNet approach. The energies from the two PESs are directly compared as are the IR spectra determined from vibrational self-consistent field/vibrational configuration interaction calculations using fitted dipole moment surfaces. In addition, tunneling splittings for the hydrogen transfer between two equivalent structures are reported using three different methods: ring polymer instanton theory, diffusion Monte Carlo simulations, and the $Q_{im}$-path method. These calculations require the evaluation of on the order of one billion energies that are widely dispersed in the 15-dimensional configurational space. The two PESs yield results for these quantities in excellent agreement with each other.

While deep learning offers powerful capabilities for scientific research, its application is often hindered by a lack of quantitative reliability. To address this, we introduce a probabilistic denoising framework that simultaneously extracts denoised signals and element-wise predictive uncertainties from noisy data. We demonstrate this approach on three-dimensional angle-resolved photoemission spectroscopy data, showing that the model reliably recovers the spectral features of a cuprate superconductor from Poisson-distributed noise with an average count of only 0.02 electrons per voxel. Crucially, we show that these predicted uncertainties can be propagated into subsequent superconducting gap analyses, enabling quantitative parameter extraction with scientifically meaningful error bars. Furthermore, we validate the broad applicability of our approach by successfully extending it to two-dimensional x-ray diffraction data. Ultimately, this approach establishes uncertainty-aware deep learning not merely as a visualization tool, but as a rigorous framework for scientific data analysis.

Foundation models build an effective representation of data that can be deployed on diverse downstream tasks. Previous research showed that the OmniLearned, a foundation model for collider physics, could significantly advance discovery potential across collider experiments. In this paper, we show that foundation models trained on collider data can transfer to cosmology. Specifically, we improve the prediction of cosmological parameters and halo and galaxy velocities in different datasets from CosmoBench. To the best of our knowledge, this is the first time a collider physics model has been shown to generalize across scientific fields.

Using a statistical mechanics framework, we investigate the parameter space of transformer models trained on protein sequence data. We sample the loss landscape at varying temperatures using Langevin dynamics to characterize the low-loss manifold and to understand the mechanisms underlying transformers’ superior performance in protein structure prediction. We find that, at variance with networks not based on the attention mechanism, the lack of a first-order-like transition in the loss of the transformer produces a range of intermediate temperatures with good learning properties; this is true for both synthetic and natural protein sequences. We also show that the parameters of most layers are highly conserved at these temperatures if the dimension of the embedding is optimal, and we provide an operative way to find this dimension. Additionally, we show that the attention matrix is more predictive of the contact maps of the protein at higher temperatures and for higher dimensions of the embedding than those optimal for learning. Finally, we showed that the models sampled at intermediate temperatures can predict the free-energy variation upon mutation, better than models obtained through standard optimization techniques.

The next generation of particle physics experiments will face an era of challenges in data acquisition, due to unprecedented data rates and volumes along with extreme environments and operational constraints. Harnessing these data for scientific discovery demands real-time inference and decision-making, intelligent data reduction, and efficient processing architectures beyond current capabilities. Crucial to the success of this experimental paradigm are several emerging technologies, such as artificial intelligence and machine learning, silicon microelectronics, and the advent of quantum algorithms and processing. Their intersection includes areas of research such as low-power and low-latency devices for edge computing, heterogeneous accelerator systems, reconfigurable hardware, codesign and synthesis strategies, readout for cryogenic or high-radiation environments, and analog computing. This white paper presents a community-driven vision to identify and prioritize research and development opportunities in hardware-based ML systems and corresponding physics applications, contributing toward a successful transition to the data frontier of fundamental science.

Extracting interpretable mathematical models from complex dynamical systems is difficult, especially for chaotic dynamics observed with noisy experimental data. We present PEM-UDE, a method that combines prediction-error methodology with universal differential equations to discover governing equations from limited, noise-corrupted observations. Prediction-error feedback smooths the chaotic optimization problem; for noise-free data generated within the model class, it preserves the data-consistent zero-loss set, whereas noise and model misspecification introduce a gain-dependent stability-bias trade-off. Preservation of the zero-loss set is not a guarantee of unique structural identifiability. We test the method on two benchmark chaotic systems—the Rössler attractor and a real electrical circuit—and recover the correct functional forms even when one observed dimension contains noise of 5 times the signal magnitude. The method also accepts prior knowledge of the system as an initial functional form, which we use to learn neural circuit equations that account for sparse connectivity, a feature missing from conventional neural mass models. Applied to a population of Izhikevich neurons, PEM-UDE yields a multiscale neural mass model that ties single-neuron parameters to macroscopic network dynamics and predicts a relationship between connection density, dominant oscillation frequency, and synchrony. We test these predictions against three intracranial recording datasets from rat and human cortices. For the neuroscience application, the learned equations are a reduced-order closure for a specified simulated Izhikevich network family; the experimental recordings provide an indirect consistency check of predicted frequency and synchrony trends, not a direct fit of the equations to recordings.

The recent development of Neural Quantum States (NQS) has established them as one of the most accurate methods for studying strongly correlated many-fermion systems, outperforming existing many-body approaches for large systems. However, NQS calculations are currently highly resource intensive. In this work, we introduce an improved Pareto frontier of efficiency and accuracy for NQS in the simulation of strongly correlated lattice fermions. This frontier is defined by two complementary backflow-related architectures: the Sparse Convolutional Ansatz for Lattice Electrons (SCALE), which achieves high computational efficiency, and the Accurate Convolutional ansatz for lattice Electrons (ACE), which delivers high variational accuracy, based on benchmark results on the Hubbard model and the $t\text{–}J$ model on large lattice sizes. SCALE utilizes a tailored convolutional design to enable an efficient local update based on the low-rank update of determinants. This structure reduces the computational scaling from $\mathcal{O}\left(N^4\right)$ to $\mathcal{O}\left(N^3\right)$ in backflow methods, resulting in a practical speedup exceeding $40\times$ in our test cases while retaining accurate variational accuracy. As an application, we study the $1/8$-doped pure Hubbard model up to size $32\times 32$ and find no significant energy difference between the horizontal and vertical unit-filled stripe states, in contrast to the half-filled stripe state when next-nearest-neighbor hoppings are included. ACE, conversely, employs a deep convolutional stack to maximize expressive power, achieving high accuracy on large systems. We perform extensive benchmark calculations on the Hubbard and $t\text{–}J$ models with the two architectures. On the challenging repulsive Hubbard model, SCALE achieves variational energies competitive with leading methods at a fraction of the computational cost. Meanwhile, ACE achieves lower variational energies than previously reported results while requiring only one-sixth of the run-time for system size $16\times 4$. These NQS approaches provide scalable, affordable, and accurate tools for exploring the rich physics of strongly correlated fermionic systems.

We present Denario, an Artificial Intelligence (AI) multiagent system designed to serve as a scientific research assistant. Denario can perform many different tasks, such as generating ideas, checking the literature, developing research plans, writing and executing code, making plots, and drafting a scientific paper. The system has a modular architecture, allowing it to handle specific tasks, such as generating an idea or carrying out near end-to-end scientific analysis using Cmbagent as a deep-research backend (end-to-end except for the novelty assessment, which is intentionally left for the user to inspect). In this work, we describe in detail Denario and its modules, and illustrate its capabilities by presenting multiple AI-generated papers generated by it in many different scientific disciplines such as astrophysics, biology, biophysics, biomedical informatics, chemistry, material science, mathematical physics, medicine, neuroscience, and planetary science. Denario can also produce work that combines ideas from different disciplines, which we illustrate with a single ($N=1$) example: a paper that applies methods from quantum physics and machine learning to astrophysical data. We report the evaluations performed on these papers by domain experts, who provided both numerical scores and reviewlike feedback. These papers were scored by domain experts on our team; we report these scores as an indicative self-assessment rather than an independent evaluation, and we flag the resulting conflict of interest explicitly. Most drafts were rated above the midpoint of a 0–10 scale, with several rated 8–9, indicating that the generated papers reach the level a good undergraduate or early graduate student could achieve after weeks or a few months of work. They should be regarded as first drafts that require expert development, not publication-ready papers. We also document an important failure mode in which, when the input data cannot be read, the system may silently generate synthetic data, which we flag explicitly as a key limitation. We then highlight the strengths, weaknesses, and limitations of the current system. Finally, we discuss the ethical implications of AI-driven research and reflect on how such technology relates to the philosophy of science. We publicly release the code at https://github.com/AstroPilot-AI/Denario. A Denario demo can also be run directly on the web at https://huggingface.co/spaces/astropilot-ai/Denario, and the full app will be deployed on the cloud.

Existing approaches to explicit data-driven discovery of nonlinear dynamical systems face a curse of dimensionality because candidate libraries grow combinatorially with system dimension. This challenge is particularly severe for high-dimensional systems arising from spatially discretized fields and large complex networks. We develop sparse regression embedded interpretable network (SREINet), a machine-learning framework that embeds sparse regression in an interpretable neural network and uses a sparsity-promoting periodic pruning scheme. Across six paradigmatic systems, SREINet accurately identifies velocity fields with more than 100 dimensions and extrapolates coherent structures from untrained data. Head-to-head benchmarks against representative equation-discovery and predictive methods on clean and noisy systems show that SREINet combines scalable computation with accurate, parsimonious equation recovery, particularly in high-dimensional settings. We further validate the framework using empirical data from a triple-pendulum experiment. These results demonstrate a practical route toward interpretable model discovery for large-scale nonlinear systems and suggest that the framework can be extended to systems with thousands of dimensions.

Tip-Enhanced Raman Spectroscopy (TERS) provides nanoscale chemical fingerprints alongside high-resolution topographic mapping of molecules, offering a powerful tool for materials discovery. However, TERS image datasets are challenging to interpret and typically demand time-consuming, computationally intensive quantum-chemistry calculations. To overcome this problem, we present an encoder-decoder model trained and evaluated on simulated TERS images of planar molecules, enabling direct prediction of molecular structures from spectral simulated data with high accuracy. Our approach demonstrates the feasibility of automating molecular structure identification from TERS images, bypassing traditional manual analysis. These findings provide a foundation for extending machine learning methods to experimental TERS datasets, potentially accelerating molecular discovery by integrating nanoscale spectroscopy with automated computational analysis.

Backpropagation learning algorithm, the workhorse of modern artificial intelligence, is notoriously difficult to implement in physical neural networks. Equilibrium propagation (EP) is an alternative with comparable efficiency and strong potential for in situ training. We extend EP learning to both discrete and continuous complex-valued wave systems. In contrast to previous EP implementations, our scheme is valid in the weakly dissipative regime and readily applicable to a wide range of physical settings, even without well-defined nodes, where trainable internode connections can be replaced by trainable local potential. We test the method in driven-dissipative exciton-polariton condensates governed by generalized Gross-Pitaevskii dynamics. Numerical studies on standard benchmarks, including a simple logical task and handwritten-digit recognition, demonstrate stable convergence, establishing a practical route to in situ learning in physical systems in which system control is restricted to local parameters.