arXiv subjects

Jules Tilly

Publications and source records attributed to Jules Tilly.

17 recordsLinked to original sources

arXiv2026

ChemReporter: A Framework for Curating and Exporting Large-Scale Chemical Datasets for MLIP Training

Training set quality and diversity are key determinants of the reliability of machine learning interatomic potentials (MLIPs), yet using massive datasets in full is often impractical and redundant, making intelligent data selection essential. A major bottleneck, however, is the lack of infrastructure for uniformly accessing, curating, and subsampling heterogeneous large-scale chemical datasets, which differ widely in structure, metadata, and file format. We address this gap with ChemReporter, a modular, method-agnostic framework that converts arbitrary molecular and materials datasets into a unified, queryable representation and exports the results directly into MLIP-ready training data. ChemReporter operates in three decoupled stages: processing, which parses raw datasets into a partitioned Apache Parquet repository enriched with structural, physical, and chemical metadata; querying, which filters and samples this repository via a CLI or Python API using arbitrary selection criteria, from simple physical constraints to custom, user-defined strategies; and exporting, which streams the selected subset into an HDF5 file ready for direct use in modern MLIP training frameworks. Throughout this process, every exported data point remains traceable to its original source entry, and dataset exports can be reliably reproduced given the same configuration and query database version. Because data is stored in a queryable, disk-backed format, ChemReporter can process datasets far larger than available memory, allowing it to scale to billion-structure datasets on standard compute infrastructure. ChemReporter is available on GitHub and PyPI under the Apache License 2.0.

physics.chem-ph↗

arXiv2026

Machine Learning Interatomic Potentials: Advancing Open-Source Software for Efficient and Scalable Molecular Simulation

Machine learning interatomic potentials (MLIPs) enable atomistic simulations with near ab initio accuracy at significantly reduced computational cost, but their broader adoption is often limited by fragmented tooling, limited scalability, and inflexible software design. We present mlip v2, a new generation of the mlip library that advances efficient and scalable molecular simulation through a unified and extensible framework. The new release features a targeted API redesign with improved modularity and control, enabling flexible customization of training, data processing, and simulation workflows. It further integrates a new high-performance backend for equivariant operations, e3j, significantly accelerating model inference and simulations. In addition, the framework introduces a range of entirely new capabilities, including the eSEN architecture with a Mixture-of-Experts formulation for scalable training on large and diverse datasets, improved handling of electrostatics through more physically grounded charge modeling and long-range interaction treatment, and advanced simulation features such as NPT ensembles and nudged elastic band methods. Together, these extensions significantly broaden the scope of MLIP applications, enabling efficient modeling of complex, reactive, and out-of-equilibrium systems, and bridging the gap between ML research and practical molecular simulation applications. The library is available on GitHub and on PyPI under the Apache license 2.0.

physics.chem-ph↗

arXiv2026

Force-Aware Neural Tangent Kernels for Scalable and Robust Active Learning of MLIPs

Active learning for machine-learning interatomic potentials (MLIPs) must address several challenges to be practical: scaling to large candidate pools, leveraging energy-force supervision, and maintaining robustness when candidate pools are biased relative to the target distribution. In this work, we jointly address these challenges. We first introduce a linearly scaling acquisition framework based on chunked feature-space posterior-variance shortlisting. By avoiding materialisation of the candidate and train set kernels, this approach enables screening of ~200k structures within hours and applies broadly to acquisition strategies that score candidates based on molecular similarity metrics. We then extend the Neural Tangent Kernel (NTK) to a force-aware setting via mixed parameter-coordinate derivatives, yielding a force NTK and a joint energy-force NTK that provide natural similarity metrics for vector-field prediction. We demonstrate the effectiveness of the joint energy-force NTK on the OC20 dataset, where force-aware acquisition is crucial: it achieves the lowest energy and force MAE and RMSE across all metrics and distribution splits. Across T1x, PMechDB, and RGD benchmarks, our force NTK methods remain competitive with established baselines while being significantly more efficient than committee-based approaches. Under a controlled candidate-pool shift case study on T1x, acquisition based on pretrained MLIP embeddings and NTKs remains robust, whereas committee-based methods exhibit higher variance. Overall, these results show that a single pretrained MLIP can enable scalable, force-aware, and distribution-robust active learning for foundation-model fine-tuning.

cs.LG↗

arXiv2026

Pretrained Model Representations as Acquisition Signals for Active Learning of MLIPs

Training machine learning interatomic potentials (MLIPs) for reactive chemistry is often bottlenecked by the high cost of quantum chemical labels and the scarcity of transition state configurations in candidate pools. Active learning (AL) can mitigate these costs, but its effectiveness hinges on the acquisition rule. We investigate whether the latent space of a pretrained MLIP already contains the information necessary for effective acquisition, eliminating the need for auxiliary uncertainty heads, Bayesian training and fine-tuning, or committee ensembles. We introduce two acquisition signals derived directly from a pretrained MACE potential: a finite-width neural tangent kernel (NTK) and an activation kernel built from hidden latent space features. On reactive-chemistry benchmarks, both kernels consistently outperform fixed-descriptor baselines, committee disagreement, and random acquisition, reducing the data required to reach performance targets by an average of 38% for energy error and 28% for force error. We further show that the pretrained model induces similarity spaces that preserve chemically meaningful structure and provide more reliable residual uncertainty estimates than randomly initialised or fixed-descriptor-based kernels. Our results suggest that pretraining aligns latent-space geometry with model error, yielding a practical and sufficient acquisition signal for reactive MLIP fine-tuning.

cs.LG↗

arXiv2026

Integral Formulas for Vector Signal Tensor Products

We derive integral formulas that simplify the Vector Signal Tensor Product recently introduced by Xie et al., which generalizes the Gaunt tensor product to anti-symmetric couplings. In particular, we obtain explicit closed-form expressions for the anti-symmetric analogues of the Gaunt coefficients. This enables us to simulate the Clebsch-Gordan tensor product using a single Vector Signal Tensor Product, yielding up to a $9\times$ reduction in the required tensor product evaluations. Our results enable efficient and practical implementations of the Vector Signal Tensor Product, paving the way for applications of this generalization of Gaunt Tensor Products in $\mathrm{SO}(3)$-equivariant neural networks. Moreover, we discuss how the Gaunt and the Vector Signal Tensor Products allow to control the expressivity-runtime tradeoff associated with the usual Clebsch-Gordan Tensor Products. Finally, we investigate low rank decompositions of the normalizations of the considered tensor products in view of their use in equivariant neural networks.

cs.LG↗

arXiv2025

MLIPAudit: A benchmarking tool for Machine Learned Interatomic Potentials

Machine-learned interatomic potentials (MLIPs) promise to significantly advance atomistic simulations by delivering quantum-level accuracy for large molecular systems at a fraction of the computational cost of traditional electronic structure methods. While model hubs and categorisation efforts have emerged in recent years, it remains difficult to consistently discover, compare, and apply these models across diverse scenarios. The field still lacks a standardised and comprehensive framework for evaluating MLIP performance. We introduce MLIPAudit, an open, curated and modular benchmarking suite designed to assess the accuracy of MLIP models across a variety of application tasks. MLIPAudit offers a diverse collection of benchmark systems, including small organic compounds, molecular liquids, proteins and flexible peptides, along with pre-computed results for a range of pre-trained and published models. MLIPAudit also provides tools for users to evaluate their models using the same standardised pipeline. A continuously updated leaderboard tracks performance across benchmarks, enabling direct comparison on downstream tasks. By providing a unified, transparent reference framework for model validation and comparison, MLIPAudit aims to foster reproducibility, transparency, and community-driven progress in the development of MLIPs for complex molecular systems. In order to illustrate the use of the library, we present some benchmarks run on a series of internal models, along with publicly available ones (UMA-Small, MACE-OFF, MACE-MP). The library is available on GitHub at https://github.com/instadeepai/mlipaudit, on PyPI at https://pypi.org/project/mlipaudit/ under the Apache License 2.0, and the leaderboard can be accessed on HuggingFace at https://huggingface.co/spaces/InstaDeepAI/mlipaudit-leaderboard.

physics.chem-ph↗

arXiv2025

Machine Learning Interatomic Potentials: library for efficient training, model development and simulation of molecular systems

Machine Learning Interatomic Potentials (MLIP) are a novel in silico approach for molecular property prediction, creating an alternative to disrupt the accuracy/speed trade-off of empirical force fields and density functional theory (DFT). In this white paper, we present our MLIP library which was created with two core aims: (1) provide to industry experts without machine learning background a user-friendly and computationally efficient set of tools to experiment with MLIP models, (2) provide machine learning developers a framework to develop novel approaches fully integrated with molecular dynamics tools. The library includes in this release three model architectures (MACE, NequIP, and ViSNet), and two molecular dynamics (MD) wrappers (ASE, and JAX-MD), along with a set of pre-trained organics models. The seamless integration with JAX-MD, in particular, facilitates highly efficient MD simulations, bringing MLIP models significantly closer to industrial application. The library is available on GitHub and on PyPI under the Apache license 2.0.

physics.chem-ph↗

arXiv2024

BoostMD: Accelerating molecular sampling by leveraging ML force field features from previous time-steps

Simulating atomic-scale processes, such as protein dynamics and catalytic reactions, is crucial for advancements in biology, chemistry, and materials science. Machine learning force fields (MLFFs) have emerged as powerful tools that achieve near quantum mechanical accuracy, with promising generalization capabilities. However, their practical use is often limited by long inference times compared to classical force fields, especially when running extensive molecular dynamics (MD) simulations required for many biological applications. In this study, we introduce BoostMD, a surrogate model architecture designed to accelerate MD simulations. BoostMD leverages node features computed at previous time steps to predict energies and forces based on positional changes. This approach reduces the complexity of the learning task, allowing BoostMD to be both smaller and significantly faster than conventional MLFFs. During simulations, the computationally intensive reference MLFF is evaluated only every $N$ steps, while the lightweight BoostMD model handles the intermediate steps at a fraction of the computational cost. Our experiments demonstrate that BoostMD achieves an eight-fold speedup compared to the reference model and generalizes to unseen dipeptides. Furthermore, we find that BoostMD accurately samples the ground-truth Boltzmann distribution when running molecular dynamics. By combining efficient feature reuse with a streamlined architecture, BoostMD offers a robust solution for conducting large-scale, long-timescale molecular simulations, making high-accuracy ML-driven modeling more accessible and practical.

physics.chem-ph↗

arXiv2024

Unified Framework for Matchgate Classical Shadows

Estimating quantum fermionic properties is a computationally difficult yet crucial task for the study of electronic systems. Recent developments have begun to address this challenge by introducing classical shadows protocols relying on sampling of Fermionic Gaussian Unitaries (FGUs): a class of transformations in fermionic space which can be conveniently mapped to matchgates circuits. The different protocols proposed in the literature use different sub-ensembles of the orthogonal group $O(2n)$ to which FGUs can be associated. We propose an approach that unifies these different protocols, proving their equivalence, and deriving from it an optimal sampling scheme. We begin by demonstrating that the first three moments of the FGU ensemble associated with $SO(2n)$ and of its intersection with the Clifford group are equal, generalizing a result known for $O(2n)$ and addressing a question raised in previous works. Building on this proof, we establish the equivalence between the shadows protocols resulting from FGU ensembles analyzed in the literature. Finally, from our results, we propose a sampling scheme for a small sub-ensemble of matchgates circuits that is optimal in terms of number of gates and that inherits the performances guarantees of the previous ensembles.

quant-ph↗

arXiv2023

A foundation model for atomistic materials chemistry

Atomistic simulations of matter, especially those that leverage first-principles (ab initio) electronic structure theory, provide a microscopic view of the world, underpinning much of our understanding of chemistry and materials science. Over the last decade or so, machine-learned force fields have transformed atomistic modeling by enabling simulations of ab initio quality over unprecedented time and length scales. However, early ML force fields have largely been limited by: (i) the substantial computational and human effort of developing and validating potentials for each particular system of interest; and (ii) a general lack of transferability from one chemical system to the next. Here we show that it is possible to create a general-purpose atomistic ML model, trained on a public dataset of moderate size, that is capable of running stable molecular dynamics for a wide range of molecules and materials. We demonstrate the power of the MACE-MP-0 model - and its qualitative and at times quantitative accuracy - on a diverse set of problems in the physical sciences, including properties of solids, liquids, gases, chemical reactions, interfaces and even the dynamics of a small protein. The model can be applied out of the box as a starting or "foundation" model for any atomistic system of interest and, when desired, can be fine-tuned on just a handful of application-specific data points to reach ab initio accuracy. Establishing that a stable force-field model can cover almost all materials changes atomistic modeling in a fundamental way: experienced users get reliable results much faster, and beginners face a lower barrier to entry. Foundation models thus represent a step towards democratising the revolution in atomic-scale modeling that has been brought about by ML force fields.

physics.chem-ph↗

arXiv2023

Encoding optimization for quantum machine learning demonstrated on a superconducting transmon qutrit

Qutrits, three-level quantum systems, have the advantage of potentially requiring fewer components than the typically used two-level qubits to construct equivalent quantum circuits. This work investigates the potential of qutrit parametric circuits in machine learning classification applications. We propose and evaluate different data-encoding schemes for qutrits, and find that the classification accuracy varies significantly depending on the used encoding. We therefore propose a training method for encoding optimization that allows to consistently achieve high classification accuracy. Our theoretical analysis and numerical simulations indicate that the qutrit classifier can achieve high classification accuracy using fewer components than a comparable qubit system. We showcase the qutrit classification using the optimized encoding method on superconducting transmon qutrits, demonstrating the practicality of the proposed method on noisy hardware. Our work demonstrates high-precision ternary classification using fewer circuit elements, establishing qutrit parametric quantum circuits as a viable and efficient tool for quantum machine learning applications.

quant-ph↗

arXiv2021

The Variational Quantum Eigensolver: a review of methods and best practices

The variational quantum eigensolver (or VQE) uses the variational principle to compute the ground state energy of a Hamiltonian, a problem that is central to quantum chemistry and condensed matter physics. Conventional computing methods are constrained in their accuracy due to the computational limits. The VQE may be used to model complex wavefunctions in polynomial time, making it one of the most promising near-term applications for quantum computing. Finding a path to navigate the relevant literature has rapidly become an overwhelming task, with many methods promising to improve different parts of the algorithm. Despite strong theoretical underpinnings suggesting excellent scaling of individual VQE components, studies have pointed out that their various pre-factors could be too large to reach a quantum computing advantage over conventional methods. This review aims to provide an overview of the progress that has been made on the different parts of the algorithm. All the different components of the algorithm are reviewed in detail including representation of Hamiltonians and wavefunctions on a quantum computer, the optimization process, the post-processing mitigation of errors, and best practices are suggested. We identify four main areas of future research:(1) optimal measurement schemes for reduction of circuit repetitions; (2) large scale parallelization across many quantum computers;(3) ways to overcome the potential appearance of vanishing gradients in the optimization process, and how the number of iterations required for the optimization scales with system size; (4) the extent to which VQE suffers for quantum noise, and whether this noise can be mitigated. The answers to these open research questions will determine the routes for the VQE to achieve quantum advantage as the quantum computing hardware scales up and as the noise levels are reduced.

quant-ph↗

arXiv2021

Improving resilience of the Quantum Gravity Induced Entanglement of Masses (QGEM) to decoherence using 3 superpositions

Recently a protocol called quantum gravity induced entanglement of masses (QGEM) that aims to test the quantum nature of gravity using the entanglement of 2 qubits was proposed. The entanglement can arise only if the force between the two spatially superposed masses is occurring via the exchange of a mediating virtual graviton. In this paper, we examine a possible improvement of the QGEM setup by introducing a third mass with an embedded qubit, so that there are now 3 qubits to witness the gravitationally generated entanglement. We compare the entanglement generation for different experimental setups with 2 and 3 qubits and find that a 3-qubit setup where the superpositions are parallel to each other leads to the highest rate of entanglement generation within $\tau = 5 $ s. We will show that the 3-qubit setup is more resilient to the higher rate of decoherence. The entanglement can be detected experimentally for the 2-qubit setup if the decoherence rate $\gamma$ is $\gamma < 0.11 $ Hz compared to $\gamma < 0.16 $ Hz for the 3-qubit setup. However, the introduction of an extra qubit means that more measurements are required to characterize entanglement in an experiment. We conduct experimental simulations and estimate that the 3-qubit setup would allow detecting the entanglement in the QGEM protocol at a $99.9\%$ certainty with $O(10^4)-O(10^5)$ measurements when $\gamma \in [0.1,0.15] $ Hz. Furthermore, we find that the number of needed measurements can be reduced to $O(10^3)-O(10^5)$ if the measurement schedule is optimised using joint Pauli basis measurements. For $\gamma > 0.06 $ Hz the 3-qubit setup is favourable compared to the 2-qubit setup in terms of the minimum number of measurements needed to characterize the entanglement. Thus, the proposed setup here provides a promising new avenue for implementing the QGEM experiment.

quant-ph↗

arXiv2021

A variational quantum eigensolver for dynamic correlation functions

Recent practical approaches for the use of current generation noisy quantum devices in the simulation of quantum many-body problems have been dominated by the use of a variational quantum eigensolver (VQE). These coupled quantum-classical algorithms leverage the ability to perform many repeated measurements to avoid the currently prohibitive gate depths often required for exact quantum algorithms, with the restriction of a parameterized circuit to describe the states of interest. In this work, we show how the calculation of zero-temperature dynamic correlation functions defining the linear response characteristics of quantum systems can also be recast into a modified VQE algorithm, which can be incorporated into the current variational quantum infrastructure. This allows for these important physical expectation values describing the dynamics of the system to be directly converged on the frequency axis, and they approach exactness over all frequencies as the flexibility of the parameterization increases. The frequency resolution hence does not explicitly scale with gate depth, which is approximately twice as deep as a ground state VQE. We apply the method to compute the single-particle Green's function of ab initio dihydrogen and lithium hydride molecules, and demonstrate the use of a practical active space embedding approach to extend to larger systems. While currently limited by the fidelity of two-qubit gates, whose number is increased compared to the ground state algorithm on current devices, we believe the approach shows potential for the extraction of frequency dynamics of correlated systems on near-term quantum processors.

quant-ph↗

arXiv2021

Reduced Density Matrix Sampling: Self-consistent Embedding and Multiscale Electronic Structure on Current Generation Quantum Computers

We investigate fully self-consistent multiscale quantum-classical algorithms on current generation superconducting quantum computers, in a unified approach to tackle the correlated electronic structure of large systems in both quantum chemistry and condensed matter physics. In both of these contexts, a strongly correlated quantum region of the extended system is isolated and self-consistently coupled to its environment via the sampling of reduced density matrices. We analyze the viability of current generation quantum devices to provide the required fidelity of these objects for a robust and efficient optimization of this subspace. We show that with a simple error mitigation strategy and optimization of compact tensor product bases to minimize the number of terms to sample, these self-consistent algorithms are indeed highly robust, even in the presence of significant noises on quantum hardware. Furthermore, we demonstrate the use of these density matrices for the sampling of non-energetic properties, including dipole moments and Fermi liquid parameters in condensed phase systems, achieving a reliable accuracy with sparse sampling. It appears that uncertainties derived from the iterative optimization of these subspaces is smaller than variances in the energy for a single subspace optimization with current quantum hardware. This boosts the prospect for routine self-consistency to improve the choice of correlated subspaces in hybrid quantum-classical approaches to electronic structure for large systems in this multiscale fashion.

quant-ph↗

arXiv2021

Qudits for Witnessing Quantum Gravity Induced Entanglement of Masses Under Decoherence

Recently a theoretical and an experimental protocol known as quantum gravity induced entanglement of masses (QGEM) has been proposed to test the quantum nature of gravity using two mesoscopic masses each placed in a superposition of two locations. If, after eliminating all non-gravitational interactions between them, the particles become entangled, one can conclude that the gravitational potential is induced via a quantum mediator, i.e. a virtual graviton. In this paper, we examine a range of different experimental set-ups, considering different geometries and the number of spatially superposed states taken, in order to determine which would generate entanglement faster. We conclude that without decoherence, and given a maximum distance $\Delta x$ between any two spatial states of a superposition, a set of two qubits placed in spatial superposition parallel to one another will outperform all other models given realistic experimental parameters. Furthermore, when a sufficiently high decoherence rate is introduced, multi-component superpositions can outperform the two-qubit set-up. This is further verified with an experimental simulation, showing that $O(10^3)$ measurements are required to reject the no entanglement hypothesis with a parallel qubits set-up without decoherence at a 99.9$\%$ confidence level. The number of measurements increases when decoherence is introduced. When the decoherence rate reaches $0.125$~Hz, 6-dimensional qudits are required as the two-qubit system entanglement cannot be witnessed anymore. However, in this case, $O(10^6)$ measurements will be required. One can group the witness operators to measure in order to reduce the number of measurements (up to ten-fold). However, this may be challenging to implement experimentally.

quant-ph↗

arXiv2020

Computation of molecular excited states on IBM quantum computers using a discriminative variational quantum eigensolver

Solving for molecular excited states remains one of the key challenges of modern quantum chemistry. Traditional methods are constrained by existing computational capabilities, limiting the complexity of the molecules that can be studied or the accuracy of the results that can be obtained. Several quantum computing methods have been suggested to address this limitation. However, these typically have hardware requirements which may not be achieved in the near term. We propose a variational quantum machine learning based method to determine molecular excited states aiming at being as resilient as possible to the defects of early Noisy Intermediate Scale Quantum (NISQ) computers and demonstrate an implementation for H2 on IBMQ. Our method uses a combination of two parametrized quantum circuits, working in tandem, combined with a Variational Quantum Eigensolver (VQE) to iteratively find the eigenstates of a molecular Hamiltonian.

quant-ph↗