# MACE

Published articles for MACE.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Compact SO(3) Equivariant Atomistic Foundation Models via Structural Pruning

DevFeed: [Compact SO(3) Equivariant Atomistic Foundation Models via Structural Pruning](<https://devfeed.tech/articles/compact-so-3-equivariant-atomistic-foundation-models-via-structural-pruning-123137.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.08885>)

Author: Chen Wang, Siyu Hu, Guangming Tan, Weile Jia

Published: 2026-11-09T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.md>), [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>), [quantization](<https://devfeed.tech/topics/quantization.md>)

Tags: [checkpoint](<https://devfeed.tech/tags/checkpoint.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data-efficiency](<https://devfeed.tech/tags/data-efficiency.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [dimensions](<https://devfeed.tech/tags/dimensions.md>), [distillation](<https://devfeed.tech/tags/distillation.md>), [energy](<https://devfeed.tech/tags/energy.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [foundation](<https://devfeed.tech/tags/foundation.md>), [foundation-models](<https://devfeed.tech/tags/foundation-models.md>), [graph-neural-networks](<https://devfeed.tech/tags/graph-neural-networks.md>), [inference-efficiency](<https://devfeed.tech/tags/inference-efficiency.md>), [mace](<https://devfeed.tech/tags/mace.md>), [pruning](<https://devfeed.tech/tags/pruning.md>), [quantization](<https://devfeed.tech/tags/quantization.md>)

### AI overview

The paper proposes pruning SO(3)-equivariant atomistic models along channel and order dimensions, removing or retaining each irreducible representation as a complete block to preserve equivariance. Pruning large checkpoints reduces inference cost while outperforming independently trained small models in the reported comparisons: pruned MACE-MP exceeds the official small model on 7 of 9 Matbench Discovery metrics. Compressed MACE-MP and MACE-OFF have 1.5x-4x fewer parameters and use 2.5x-4x less pre-training compute than training a small model from scratch. Across eight downstream datasets, fine-tuning the pruned model reduces energy and force errors by 70.1% and 34.4%, respectively, versus task-specific training from scratch. The method is also demonstrated on SevenNet and eSCN and can be combined with quantization and knowledge distillation.

### Source excerpt

arXiv:2605.08885v2 Announce Type: replace Abstract: SO(3) equivariant graph neural networks have become the dominant paradigm for atomistic foundation models, achieving high accuracy and data efficiency by building rotational symmetry directly into the architecture. Yet the computational cost of their higher-order tensor operations creates a tough trade-off between model accuracy and inference efficiency. In this paper, we propose a structural pruning method for SO(3) equivariant atomistic foundation models to bridge this accuracy-efficiency gap. The pruning is applied along the channel and order dimensions, with each irreducible representation kept or removed as a complete block, thereby retaining SO(3) equivariance. Starting from a large checkpoint, the pruned model substantially reduces the inference cost while retaining higher accuracy than an independently trained small model. The pruned MACE-MP model outperforms the official from-scratch trained small model on 7 of 9 metrics on the Matbench Discovery leaderboard. In terms of efficiency, compressed MACE-MP and MACE-OFF models contain 1.5$\times$ to 4$\times$ fewer parameters and require 2.5$\times$ to 4$\times$ less pre-training compute than training a small model from scratch. For downstream applications, fine-tuning the pruned model reduces energy and force errors by 70.1% and 34.4% compared to training task-specific models from scratch across eight representative downstream datasets. We demonstrate that the method generalizes to other SO(3) equivariant architectures (SevenNet, eSCN) and can be combined with quantization and knowledge distillation for further gains.

## Multi-Fidelity Training Reshapes Uncertainty Decomposition in Atomistic Models

DevFeed: [Multi-Fidelity Training Reshapes Uncertainty Decomposition in Atomistic Models](<https://devfeed.tech/articles/multi-fidelity-training-reshapes-uncertainty-decomposition-in-atomistic-models-103048.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.35560>)

Author: Matteo Tolloso, Davide Bacciu, Luigi Bonati

Published: 2026-10-28T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [AI Models](<https://devfeed.tech/topics/ai-models.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Predictive AI](<https://devfeed.tech/topics/predictive-ai.md>)

Tags: [acquisition](<https://devfeed.tech/tags/acquisition.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [deep](<https://devfeed.tech/tags/deep.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [energy](<https://devfeed.tech/tags/energy.md>), [error](<https://devfeed.tech/tags/error.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [mace](<https://devfeed.tech/tags/mace.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [pretraining](<https://devfeed.tech/tags/pretraining.md>), [ranking](<https://devfeed.tech/tags/ranking.md>), [training](<https://devfeed.tech/tags/training.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [variance](<https://devfeed.tech/tags/variance.md>)

### AI overview

Using deep ensembles of mean-variance MACE models trained on ANI-1x, the study compares high-fidelity training from scratch with low-fidelity pretraining followed by fine-tuning. Pretraining reduces prediction error and improves uncertainty-based ranking on unseen molecular systems and higher-energy configurations, but predictive variances remain poorly calibrated, especially under energy extrapolation. Training analysis attributes most calibration decline to high-fidelity optimization, where learned aleatoric variance becomes overconfident on held-out data. Multi-fidelity fine-tuning also yields a more effective uncertainty-based acquisition function. The results indicate that uncertainty ranking and calibration measure distinct properties and should be evaluated separately.

### Source excerpt

arXiv:2610.35560v2 Announce Type: replace-cross Abstract: Multi-fidelity training, which combines abundant low-fidelity data with scarce high-fidelity labels, is increasingly common in atomistic modeling, but its effect on predictive uncertainty remains largely unexplored. We investigate this question using deep ensembles of mean-variance MACE models trained on ANI-1x, contrasting high-fidelity training from scratch with low-fidelity pretraining followed by fine-tuning. We characterize aleatoric, epistemic, and total uncertainty throughout training and under two distribution shifts: unseen molecular systems and higher-energy configurations. We find that pretraining reduces prediction error and improves uncertainty-based ranking across all test sets, but does not yield well-calibrated predictive variances; calibration deteriorates more severely under energy extrapolation. Training-dynamics analysis localizes most of this calibration degradation to high-fidelity optimization, during which the learned aleatoric variance becomes overconfident on held-out data. Moreover, multi-fidelity fine-tuning produces a more effective uncertainty-based acquisition function than high-fidelity training from scratch. These findings show that uncertainty ranking and calibration capture distinct properties and must be assessed separately in multi-fidelity atomistic models, especially under extrapolation.

## SoLiD26: A First Principles Solid-Liquid Interface Dataset for Machine-learned Interatomic Potentials

DevFeed: [SoLiD26: A First Principles Solid-Liquid Interface Dataset for Machine-learned Interatomic Potentials](<https://devfeed.tech/articles/solid26-a-first-principles-solid-liquid-interface-dataset-for-machine-learned-interatomic-potentials-94757.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.28013>)

Author: Jonas Busk, Emil J. P. Frost, Yogeshwaran Krishnan, Henrik H. Kristoffersen, August E. G. Mikkelsen, Xueping Qin, Xin Yang, Heine A. Hansen, Arghya Bhowmik, Tejs Vegge

Published: 2026-10-22T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Dataset development](<https://devfeed.tech/topics/dataset-development.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [Molecular Dynamics](<https://devfeed.tech/topics/molecular-dynamics.md>)

Tags: [aimd](<https://devfeed.tech/tags/aimd.md>), [cond-mat-mtrl-sci](<https://devfeed.tech/tags/cond-mat-mtrl-sci.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [dft](<https://devfeed.tech/tags/dft.md>), [dispersion](<https://devfeed.tech/tags/dispersion.md>), [elements](<https://devfeed.tech/tags/elements.md>), [first](<https://devfeed.tech/tags/first.md>), [functional](<https://devfeed.tech/tags/functional.md>), [interface](<https://devfeed.tech/tags/interface.md>), [interfaces](<https://devfeed.tech/tags/interfaces.md>), [liquid](<https://devfeed.tech/tags/liquid.md>), [mace](<https://devfeed.tech/tags/mace.md>), [machine](<https://devfeed.tech/tags/machine.md>), [materials](<https://devfeed.tech/tags/materials.md>), [metal](<https://devfeed.tech/tags/metal.md>), [modelling](<https://devfeed.tech/tags/modelling.md>), [models](<https://devfeed.tech/tags/models.md>), [molecular-dynamics](<https://devfeed.tech/tags/molecular-dynamics.md>), [physics-chem-ph](<https://devfeed.tech/tags/physics-chem-ph.md>), [physics-comp-ph](<https://devfeed.tech/tags/physics-comp-ph.md>), [principles](<https://devfeed.tech/tags/principles.md>), [reference](<https://devfeed.tech/tags/reference.md>), [solid](<https://devfeed.tech/tags/solid.md>)

### AI overview

SoLiD26 is a dataset designed to address limited training data for machine-learned interatomic potentials at solid-liquid interfaces. It contains 15.4 million first-principles structures, with systems up to 576 atoms and 15 elements. Most structures come from ab initio molecular dynamics; examples include aqueous coinage-metal interfaces, electrode-electrolyte systems, and selected bulk references. They were calculated using VASP with the PBE functional and D3 dispersion corrections. The authors demonstrate use of the dataset with MACE models on a simple training, validation, and test split. The supplied text does not report model performance details.

### Source excerpt

arXiv:2609.28013v2 Announce Type: replace-cross Abstract: Machine-learned interatomic potentials (MLIPs) enable simulations of larger atomistic systems over longer timescales than traditional ab initio methods but require large, diverse training data. Existing datasets lack comprehensive sampling of solid-liquid interfaces, which exhibit complex phenomena such as solvent layering and surface reconstruction critical for energy and catalytic materials. Here, we present SoLiD26, a curated solid-liquid interface dataset of 15.4 million first-principles structures with up to 576 atoms and 15 chemical elements for training and evaluating MLIPs. The structures were compiled from density functional theory (DFT) calculations, with most configurations originating from ab initio molecular dynamics (AIMD) simulations. SoLiD26 includes aqueous coinage metal interfaces, electrode-electrolyte systems, and selected bulk reference structures, calculated with VASP using the PBE functional and D3 dispersion corrections. The application of SoLiD26 for training and evaluating MLIPs is demonstrated with a suite of MACE models on a simple training, validation and test split. This dataset provides a valuable resource for advancing MLIP development targeted at modelling solid-liquid interfaces in electrochemical and catalytic systems.

## Degree Bounds for Separating Polynomial and Neural Invariants for Point Clouds

DevFeed: [Degree Bounds for Separating Polynomial and Neural Invariants for Point Clouds](<https://devfeed.tech/articles/degree-bounds-for-separating-polynomial-and-neural-invariants-for-point-clouds-93648.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.26612>)

Author: Snir Hordan, Harm Derksen, Mircea Petrache, Nadav Dym

Published: 2026-10-21T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [polynomials](<https://devfeed.tech/topics/polynomials.md>)

Tags: [architectures](<https://devfeed.tech/tags/architectures.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [invariants](<https://devfeed.tech/tags/invariants.md>), [mace](<https://devfeed.tech/tags/mace.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [math-ac](<https://devfeed.tech/tags/math-ac.md>), [neural](<https://devfeed.tech/tags/neural.md>), [neural-networks](<https://devfeed.tech/tags/neural-networks.md>), [order](<https://devfeed.tech/tags/order.md>), [permutations](<https://devfeed.tech/tags/permutations.md>), [point](<https://devfeed.tech/tags/point.md>), [points](<https://devfeed.tech/tags/points.md>), [polynomials](<https://devfeed.tech/tags/polynomials.md>), [regular](<https://devfeed.tech/tags/regular.md>), [rotation](<https://devfeed.tech/tags/rotation.md>), [separate](<https://devfeed.tech/tags/separate.md>), [using](<https://devfeed.tech/tags/using.md>)

### AI overview

For point clouds of n points considered up to rotations and permutations, the paper establishes quadratic lower and upper bounds on the degree of separating invariant polynomials. It shows Tensor Field Networks and MACE can express all invariant polynomials of a given degree when their representation order is sufficient, yielding the first explicit quadratic bound on the order needed for completeness. Under local regularity assumptions, a message-passing model can instead separate point clouds using invariants whose degree depends on neighborhood size rather than total point count. Counterexamples show MACE still needs high-order irreducible representations to separate locally regular point clouds.

### Source excerpt

arXiv:2610.26612v1 Announce Type: new Abstract: Neural networks that respect the symmetries of point clouds commonly compute high-order representations of the rotation group to improve their expressivity. Such models are known to be universal, and thus able to separate any pair of point clouds, when using representations of sufficiently high order, yet no precise bound on this order, as a function of the number of points, was previously known. We establish the first quadratic lower and upper degree bounds for polynomials that separate point clouds of n points up to rotations and permutations. By showing Tensor Field Networks and MACE can express all invariant polynomials of a given degree with sufficient representation order, we additionally derive the first explicit, quadratic bound on the representation order sufficient for these architectures to be complete, replacing previous existence results with concrete guarantees. We further show that, under local regularity assumptions, point cloud separation can be achieved with a message-passing-based model using invariants whose degree depends only on the size of a local neighborhood, rather than on the size of the full point cloud. Finally, we provide counterexamples showing MACE must compute high order irreducible representations to separate locally regular point clouds.

## The ATHER workshop introduced students to TinyML through hands-on model building

DevFeed: [The ATHER workshop introduced students to TinyML through hands-on model building](<https://devfeed.tech/articles/tiny-models-big-curiosity-the-ather-tinyml-workshop-at-mace-60989.md>)

Original publisher: [Read original article](<https://www.seeedstudio.com/blog/2026/09/28/ather-tinyml-workshop-mace/>)

Author: Abhinav

Published: 2026-09-28T07:19:08Z

Content type: article

Language: en

Sources: [Seeed Studio](<https://devfeed.tech/sources/latest-news-from-seeed-studio.md>)

Topics: [tinyml](<https://devfeed.tech/topics/tinyml.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [AI Development](<https://devfeed.tech/topics/ai-development.md>), [Physical AI](<https://devfeed.tech/topics/physical-ai.md>)

Tags: [arduino](<https://devfeed.tech/tags/arduino.md>), [chip](<https://devfeed.tech/tags/chip.md>), [community](<https://devfeed.tech/tags/community.md>), [data](<https://devfeed.tech/tags/data.md>), [edge-ai](<https://devfeed.tech/tags/edge-ai.md>), [edge-impulse](<https://devfeed.tech/tags/edge-impulse.md>), [events](<https://devfeed.tech/tags/events.md>), [grove](<https://devfeed.tech/tags/grove.md>), [innovation](<https://devfeed.tech/tags/innovation.md>), [iot](<https://devfeed.tech/tags/iot.md>), [mace](<https://devfeed.tech/tags/mace.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [news](<https://devfeed.tech/tags/news.md>), [object-detection](<https://devfeed.tech/tags/object-detection.md>), [seeed-ranger](<https://devfeed.tech/tags/seeed-ranger.md>), [sensecraft](<https://devfeed.tech/tags/sensecraft.md>), [tinyml](<https://devfeed.tech/tags/tinyml.md>), [workshop](<https://devfeed.tech/tags/workshop.md>), [xiao](<https://devfeed.tech/tags/xiao.md>)

### AI overview

The ATHER full-day workshop introduced 72 students to TinyML, which runs machine-learning models directly on low-power devices without constant internet access. Participants used Seeed Studio's SenseCraft AI platform and hardware to build image classification and object detection models, then collected data, trained custom models with Edge Impulse, and deployed them to devices.

### Source excerpt

We usually picture AI as something massive that runs in the cloud. On September 12, The post Tiny Models, Big Curiosity: The ATHER TinyML Workshop at MACE appeared first on Latest News from Seeed Studio.

## Build It Real: Inside EVOKE 2026 at MACE

DevFeed: [Build It Real: Inside EVOKE 2026 at MACE](<https://devfeed.tech/articles/build-it-real-inside-evoke-2026-at-mace-60990.md>)

Original publisher: [Read original article](<https://www.seeedstudio.com/blog/2026/09/28/evoke-2026-project-expo-mace/>)

Author: Abhinav

Published: 2026-09-28T06:35:37Z

Content type: news

Language: en

Sources: [Seeed Studio](<https://devfeed.tech/sources/latest-news-from-seeed-studio.md>)

Topics: [AdventureX 2026](<https://devfeed.tech/topics/adventurex2026.md>), [Hackathon](<https://devfeed.tech/topics/hackathon.md>), [lerobot](<https://devfeed.tech/topics/lerobot.md>)

Tags: [arduino](<https://devfeed.tech/tags/arduino.md>), [community](<https://devfeed.tech/tags/community.md>), [competition](<https://devfeed.tech/tags/competition.md>), [edge-ai](<https://devfeed.tech/tags/edge-ai.md>), [embedded](<https://devfeed.tech/tags/embedded.md>), [embedded-systems](<https://devfeed.tech/tags/embedded-systems.md>), [esp32s3](<https://devfeed.tech/tags/esp32s3.md>), [events](<https://devfeed.tech/tags/events.md>), [evoke](<https://devfeed.tech/tags/evoke.md>), [grove](<https://devfeed.tech/tags/grove.md>), [hackathon](<https://devfeed.tech/tags/hackathon.md>), [innovation](<https://devfeed.tech/tags/innovation.md>), [iot](<https://devfeed.tech/tags/iot.md>), [mace](<https://devfeed.tech/tags/mace.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [news](<https://devfeed.tech/tags/news.md>), [project-expo](<https://devfeed.tech/tags/project-expo.md>), [seeed-ranger](<https://devfeed.tech/tags/seeed-ranger.md>), [tinyml](<https://devfeed.tech/tags/tinyml.md>), [xiao](<https://devfeed.tech/tags/xiao.md>)

### AI overview

EVOKE 2026, the MACE IoT Club's project expo and competition, selected 40 of more than 1,000 registrants to present working hardware prototypes. The event featured engineering projects and workshops spanning embedded systems, IoT, AI, and several applied challenge areas.

### Source excerpt

More than 1,000 students signed up. 40 teams made the final cut. And on September The post Build It Real: Inside EVOKE 2026 at MACE appeared first on Latest News from Seeed Studio.