Matlantis™

Matlantis™ is a high-speed universal atomistic simulator that accelerates materials discovery for a sustainable future.
Japanese account: @matlantis_ja

[Matlantis™ @matlantis_en]

A new case study from Sumitomo Rubber Industries (DUNLOP) maps how a rubber antidegradant reacts with ozone, and which pathway proceeds most readily. Elucidating this reaction mechanism is expected to contribute to the development of high-performance, long-life rubber materials.

Ozone degrades rubber products, lowering their physical properties and causing cracks. Ozone-scavenging antidegradants such as 6PPD protect the rubber by reacting with ozone first, turning into a range of oxidation products. Many of these products have been identified experimentally, but the reaction mechanism is not yet fully understood.

🔬 Using GRRM20 with Matlantis, the team ran an automated reaction path search (SC-AFIR) on MePPD, a model compound of 6PPD, with one ozone molecule, evaluating energies with Matlantis PFP. The search found 836 equilibrium structures and 2,106 energy maxima along the paths connecting them.

💡 Among the main pathways, cleavage of the benzene ring at the center of the molecule has the lowest activation energy (40.0 kJ/mol) and is the most exothermic (−508.8 kJ/mol), making it the most likely to proceed. This is consistent with experimental results from a previous study. Most pathways in that study were tentative; the calculations give activation energies and reaction enthalpies for each, so their likelihood can be compared.

🌱 Sumitomo Rubber Industries plans to use these findings in designing higher-performance antidegradants.

Read the full case study: https://t.co/fPvaLnCQP6

#Matlantis #ReactionMechanism #PolymerScience

[Matlantis™ @matlantis_en]

This Thursday: our webinar on crystal structure prediction in practice. In materials development, even a slim chance of success can justify another long round of synthesis, and crystal structure prediction offers a way to screen candidates computationally first.

By generating many atomic configurations and compositions and comparing them on a common energy scale, it shows which structures are energetically unfavorable. Those candidates can be set aside, and lab time redirected to more promising ones. The same approach can map how much of an added element a known structure tolerates before the phase becomes unstable.

Matlantis CSP runs these searches on Matlantis PFP, a universal machine learning interatomic potential covering 96 elements. It has found stable crystal structures not reported in the Materials Project database, in systems such as In-Li, As-V, Al-Li-Pd and La-Mo-O.

In the webinar, Kohei Shinohara (Preferred Networks), who works on the development of Matlantis CSP, covers how to put this into practice: choosing a search method, designing the search conditions, and reading the results so they connect to experiments and next steps.

📅 Thursday, October 1, 2026
🕒 7:00 AM PDT / 10:00 AM EDT / 3:00 PM BST / 4:00 PM CEST (11:00 PM JST)
💻 Online (Zoom), free to attend

Register: https://t.co/vi0N4ayNoU
#Matlantis #CrystalStructurePrediction #MLIP

[Matlantis™ @matlantis_en]

💧 The hydrogen economy will run through metal tanks and pipelines, and hydrogen makes that metal brittle.

One defense is an alumina film grown by atomic layer deposition, yet the film itself traps hydrogen: low-temperature growth leaves unreacted OH groups that lower its dielectric strength and thermal conductivity, and no instrument can see what form it takes. In npj Computational Materials, Dr. Turlo's group at Empa ran molecular dynamics in LAMMPS with the machine learning interatomic potential Matlantis PFP, generating amorphous structures by annealing defective bayerite at ALD growth temperatures instead of melt-quenching.

⚗️ The structures reproduce the measured density, composition, Al-O bond length and coordination number up to 36.4 at.% hydrogen, where melt-quenching yields H2 absent from real films. The team predicted the Al Auger parameter shifts XPS measures, reading hydrogen chemistry off them: high hydrogen content puts almost all of it into OH ligands, while at the lowest content the prediction falls short of experiment, marking the co-existence of interstitial protons and bridging O-H...O configurations.

🌱 Tying those shifts to hydrogen chemical states turns a routine XPS run into a probe of hydrogen inside a film, aiding the design of hydrogen barriers, separation membranes and fuel cells.

📄 https://t.co/WgJQ7m5csq

#Matlantis #Hydrogen

[Preferred Networks @PreferredNet]

PFN group delivering full-stack AI tech from custom chips to end-user service: PFN and PFCI will start accelerating selected calculations (for crystal structures, etc.) on the Matlantis™ materials simulator with PFN's MN-Core™ 2 AI chips in 2026.
https://t.co/8iwewL9m5u

[Matlantis™ @matlantis_en]

Interview: how the developers of Matlantis PFP built the universal machine learning interatomic potential (MLIP) that many said was impossible.

Before PFP, there was no such thing as a universal model. Quantum chemistry cost scales as the cube of the atom count, so every new material meant building a new one-off model, and simulating a few dozen atoms on a catalyst surface could take six months to two years.

So Takamoto and Chikashi Shinagawa, who led the development, look back on how that changed.

⚛️ Some of what they cover:
- The failure mode that shaped the design: a model that learns the rules from data alone can look highly accurate on paper, and then errors accumulate until the structure explodes mid-simulation
- How equivariance and higher-order tensor representations encode physical law in the network architecture itself, so the model cannot return behavior physics forbids
- Why DFT data generation, not model training, consumes the bulk of the compute: more than 3,200 GPU-years so far, with over 90% of the data still designed by hand
- Why Matlantis was deliberately built as an API-driven platform for professionals instead of a one-button tool

🔬 The future they describe is a quiet one: when you set out to develop a new material, first you buy a microscope, then you bring in Matlantis. Simulation as ordinary lab infrastructure, not news.

Read the full interview: https://t.co/cEsrO9nFte

#Matlantis #MaterialsScience #MachineLearning

[Shinnosuke Furuya @sfuruyaz]

本日開催！

Matlantis User Conference 2026で、NVIDIAのJustin Smithのトークがあります。せっかくの機会なので、ご参加される方はJustinといろいろ話してみてください。日本からもApplication Engineerが参加してますので、英語が苦手な方はその人を頼って…。笑

[Matlantis™ @matlantis_en]

⚡ Hydrogen from water electrolysis needs better OER catalysts. ENEOS Holdings screened roughly 100 million candidate structures.

Announced today: ENEOS Holdings Corporation is using Matlantis PFP, our universal machine learning interatomic potential, together with NVIDIA ALCHEMI to accelerate the discovery of new catalyst materials. Oxygen evolution reaction catalysts are critical to hydrogen production through water electrolysis, and the candidate space is far larger than any synthesis campaign can cover.

⚗️ PFP predicted properties across that space while maintaining quantum-level accuracy, and ALCHEMI supplied the accelerated computing to run the screening at scale. Approximately 100 million candidate structures were evaluated, priority candidates were identified for synthesis and experimental validation, and a discovery process that traditionally took years was completed in a few months.

🌱 "The collaboration between ENEOS, NVIDIA and Matlantis demonstrates what is possible when industry leaders bring together deep materials expertise, advanced computing and AI," said Daisuke Okanohara, CEO of Matlantis. "Together, we are helping move materials discovery beyond the limits of traditional approaches."

Read the full release: https://t.co/7tmUzF1bVv

#Matlantis #NVIDIA #ENEOS

[Matlantis™ @matlantis_en]

🧪 In development with Resonac: a Matlantis feature to let experimental chemists run simulations themselves, no programming needed.

This week we announced a joint demonstration program with Resonac Corporation on Matlantis Case Studio, a feature the two companies have been developing and testing together since October 2025. An alpha version is planned for the near future, primarily for companies that took part in the demonstration.

AI-powered atomistic simulation can run alongside experiments and speed up materials R&D, but its use has largely stayed with computational specialists. For an experimental chemist, testing a hypothesis through computation has meant waiting on a specialist.

⚙️ What Case Studio is being designed to do:
- Start from an existing computational case that matches your material and property of interest
- Set the conditions and run the calculation through dialogue with AI, combined with simple selections and inputs
- Show results summarized in natural language, with execution histories saved for comparison

🌱 "Matlantis is at a stage of evolving from a tool for computational specialists into a platform that can be used by all researchers, including experimental chemists," said Daisuke Okanohara, CEO of Matlantis. "Starting with Matlantis Case Studio, we will bring the possibilities of AI for Science to more researchers and further broaden its use."

Read the full release: https://t.co/WqnNokOruv

#Matlantis #Resonac

[Matlantis™ @matlantis_en]

You're working on a composition with no reported crystal structure. What do you actually do next?

Crystal structure prediction answers that question, but its place in a real project is rarely obvious: at which phase to apply it, and how to read what comes back.

In our next webinar, Kohei Shinohara, a researcher at Preferred Networks involved in the development of Matlantis CSP, presents crystal structure prediction not as the task of running calculations but as a process that supports research decisions. Matlantis CSP runs its searches on Matlantis PFP, a universal machine learning interatomic potential covering 96 elements, so no system-specific model needs to be trained.

⚗️ What the session covers:
- Where CSP fits in early-phase research, while the structure is still undetermined
- Choosing between a global search and a derivative structure search, based on the research objective
- How to design the search conditions
- How to post-analyze results and connect them to experiments and to subsequent steps

📅 Thursday, October 1, 2026
🕒 7:00 AM PDT / 10:00 AM EDT / 3:00 PM BST / 4:00 PM CEST (11:00 PM JST)
💻 Online (Zoom), free to attend

Reserve your seat:
https://t.co/vi0N4ayNoU
#Matlantis #CrystalStructurePrediction

[Matlantis™ @matlantis_en]

Matlantis User Conference 2026 is on September 18 in Tokyo. NVIDIA's Justin S. Smith joins the Keynote Session as special guest.

🎤 The keynote pairs Matlantis CEO Daisuke Okanohara, on how atomistic insight reshapes R&D decisions in the era of AI for Science, with Justin as special guest. Justin is Principal Developer Relations Manager for AI in chemistry and materials science at NVIDIA. He developed the ANI potential, which pioneered universal machine learning interatomic potentials for small-molecule drug design, and has since contributed to the HIP-NN and AIMNet MLIPs and to active learning methods for training MLIPs in materials science and reactive chemistry.

🔬 The user programme is the core of the conference. Four presentations from corporate and academic researchers show how Matlantis is used in materials research and what it has changed in their R&D. A panel discussion asks how AI is reshaping the co-creation of experiment, calculation, and data, moderated by science educator Yobinori Takumi. Company and university researchers show case-study posters, on display all day. Masashi Hattori of MEXT, Japan's science and technology ministry, closes on AI for materials and the knowledge value chain.

📅 10:45 to 20:00 JST, for corporate and academic Matlantis users.

Programme details: https://t.co/FGUBHjxtMY

#Matlantis #AIforScience #MLIP #MaterialsScience

[Matlantis™ @matlantis_en]

🔬 Oxide nanoparticles for catalysis, sensors and batteries get their facets set by humidity during processing, not just by the recipe.

MgO nanoparticles have been measured expanding in moist air since the 1960s, the opposite of gold and platinum as they shrink. That implies a negative surface stress, but what hydroxylation does to surface stress had never been calculated. In Scripta Materialia, Dr. Turlo's group at Empa ran molecular statics in LAMMPS with the machine learning interatomic potential Matlantis PFP, computing surface energy against strain for MgO's {100}, {110} and {111} facets from bare to fully hydroxylated.

⚗️ {100} holds a positive surface stress at every coverage; {110} and {111} stay negative. A cubic particle therefore cannot reach tension, however much water it adsorbs. It has to reconstruct. Wulff constructions from the computed surface energies show exactly that: {110} facets emerge above roughly 3.3 water molecules per nm², and {111} takes over past 5.6.

🌱 The predicted strain bounds bracket every data point from three experimental studies spanning 1966 to 2017. The authors expect the same treatment to hold for other surface ligands and other oxides, so water coverage and annealing protocol become deliberate levers for tuning active sites and interfacial chemistry.

📄 https://t.co/uJdJ1pQRVH

#Matlantis #Nanoparticles #SurfaceScience

[Matlantis™ @matlantis_en]

NVIDIA, our partner on the ALCHEMI Toolkit, has published a guide to building materials simulation workflows with AI coding agents. The short version: put the science in the prompt, and let the skills carry the API.

It rests on a systematic study: 45 pipelines generated by a coding agent across three workflows (silicon equation of state, oxygen adsorption on Cu(111), liquid-lithium self-diffusion) and five levels of prompt detail, scored on code structure and then run on H200 GPUs.

⚙️ What the guidance comes down to:
• Name the material, the phase, and the reference convention. Under-specifying it once turned a request for a lithium transport property into an argon demo.
• State the protocol. Scripts left to pick their own ensemble ran Langevin dynamics in production and reported lithium diffusion 3-5x too low: statistically clean, physically damped.
• Describe the constraint, not the implementation. Naming a pipeline construct changed none of 12 implementations, because the API patterns come from the skills, not the prompt.
• Ask for premise checks and an independent reference explicitly. Agents add neither on their own, and not one asked whether the requested property was physically well-posed.

🔎 What extra prompt detail did buy was code structure and about 4x the tokens, never accuracy. The science was right from the loosest prompt.

🌱 That third point is why we published skills rather than leaving the API to documentation. The Matlantis Skills library has been on GitHub since May, alongside our Claude Code integration.

Read the full guide: https://t.co/eTvIqiTsEL

#Matlantis #MLIP #AIAgents

[Matlantis™ @matlantis_en]

Can atomistic simulation answer a process question before the wafers run? Our semiconductor webinar is now available on demand.

Matlantis application scientists Qing-Jie Li and Joshua Young walk through where machine-learning atomistic simulation fits in semiconductor process development today, from precursor design to unit processes and interfaces. Matlantis PFP, a universal MLIP covering 96 elements, supplies the chemistry at DFT level with no system-specific model training required.

⚗️ Five worked examples:
- Reaction pathways of Co ALD precursors on Si substrates, with ligand substitution effects evaluated quantitatively
- High-throughput screening of area-selective ALD candidates, combining PFP adsorption energies with regression models and Bayesian optimization
- Dry etching of SiO₂ by HF gas at ~70,000 atoms, and force-dependent material removal in CMP-style polishing, enabled by LightPFP
- Melting behavior at AlN/Cu bonding interfaces, and TaN/Cu interfacial thermal resistance validated against experiment
- Property prediction beyond energies and forces using PFP Descriptors

🎯 Useful whether you develop precursors, own a unit process, or are building simulation capability in house.

🎥 Watch the recording (free): https://t.co/tU7mG3II9n

#Matlantis #Semiconductor

[Matlantis™ @matlantis_en]

Choosing MD simulation software isn't about picking the most famous one. Our new blog sets out six questions to ask first, then compares 14 tools against them.

It starts with the force calculation, which consumes most of the compute time. Classical MD can reach microseconds and millions of atoms. First-principles MD handles bond formation and breaking but stops around a few hundred atoms and picoseconds. MLIP-MD, driven by a machine learning interatomic potential (MLIP), now brings tens of thousands of atoms and tens of nanoseconds into range.

⚙️ Six points to check before you commit:
• Coverage of your material systems, target phenomena, and the analysis functions you need
• Reachable system size and time scale, including GPU and parallel support
• Extensibility, meaning whether you can add your own potentials or analysis
• MLIP support, which means exporting DFT training data in one type of software and embedding an external MLIP in another
• What adoption and continued use actually cost: license, hardware, documentation, support
• Expertise required, with the twist that script-driven tools are now the easier ones to pair with coding agents

📊 Compared side by side: LAMMPS, OpenMM, GROMACS, AMBER, NAMD, GENESIS, VASP, Quantum ESPRESSO, CP2K, SIESTA, OpenMX, Materials Studio, J-OCTA, Desmond.

🌱 Matlantis takes the MLIP route with a general-purpose potential, Matlantis PFP: no dedicated MLIP to train per system, no DFT dataset to assemble first, and MD run by calling PFP from ASE or LAMMPS in the browser.

Read the full piece: https://t.co/NqO3Ikk8ys

#Matlantis #MolecularDynamics #MLIP

[Matlantis™ @matlantis_en]

Phase stability from CALPHAD. Atomic detail from MLIPs. An AI agent running the simulations. Prof. Yu Zhong's webinar is now on demand.

In it, he walks through how his group at Worcester Polytechnic Institute combines thermodynamic modeling, density functional theory, and machine learning interatomic potentials into a single materials design workflow. For the atomistic layer, his group works with several MLIPs, running part of that work on Matlantis PFP.

⚗️ What the session covers:
- High-throughput CALPHAD and ab initio studies of phase stability in high-entropy alloys and oxide systems
- Machine-learning-accelerated atomistic simulations using MLIPs and the Matlantis platform
- Case studies on high-entropy perovskites, interface design in oxide systems, and active-learning simulations of doping effects
- Masgent, an AI agent for autonomous simulation pipelines

🔋 The battery section covers high-entropy cathodes, solid-state electrolytes, and sodium-ion materials.

Aimed at computational materials scientists, alloy designers working on steels, superalloys, and HEAs, and battery researchers.

🎥 Watch the recording (free): https://t.co/lAGzgBTG4w

#Matlantis #CALPHAD #AIAgents

[Matlantis™ @matlantis_en]

⚡ A fuel cell cathode catalyst that gets more active and more durable at the same time, from decorating a Pd-Pt-W alloy with iron hydroxide.

Published in ACS Applied Materials & Interfaces, work from Prof. Jeong Woo Han's group at Seoul National University incorporated iron hydroxide into a Pd-Pt-W ternary alloy core-shell. The decoration exposes more metal-hydroxide interface sites, and the small quantities of W and Fe reconstruct the surface, altering the electronic structure of the active sites and leaving defects across it. Oxygen reduction reaction (ORR) activity and durability both improve.

⚗️ To pin down what the iron hydroxide actually is under operating conditions, the team built surface Pourbaix diagrams from first-principles calculations, finding which phase stays stable inside the catalysts' ORR window of 0.75 to 0.85 V vs RHE. They relaxed the candidate clusters on the Pt shell with the machine learning interatomic potential Matlantis PFP in a genetic algorithm search, then refined them in DFT. Those interfacial sites are where ORR intermediates adsorb, modulating the OH and OOH binding free energies.

🌱 Metal and metal hydroxide interfaces are a familiar lever in electrocatalysis. Pairing the experimental synthesis with theory that resolves the structure of the decorating phase is what makes tailoring these catalysts precise.

📄 https://t.co/b2LmYKL4Mj

Note: The image is an AI-generated illustration to visualize the concept.

#Matlantis #Electrocatalysis #FuelCells

[Matlantis™ @matlantis_en]

⚖️ Matlantis PFP v9 is tied for 1st overall on MLIP Arena, and ranks 1st in three of the five benchmark tasks.

MLIP Arena (Chiang et al., NeurIPS 2025) is an open benchmark that looks past energy and force prediction errors, asking instead whether a model reproduces physically meaningful behavior in practical simulation settings: homonuclear diatomics, bulk equation of state, energy-volume scans, stability, and combustion.

🔬 The tie is with MACE-MPA, against leaderboard values as of March 1, 2026. The three tasks PFP v9 leads are diatomics, equation of state, and stability. Leading three of the five means PFP v9 comes out ahead on more tasks than any other model in a one-on-one comparison.

⚛️ We evaluated PFP v9 in its PBE calculation mode, since no public r2SCAN-based benchmark exists yet. The r2SCAN mode is where v9 improved most, now covering all 96 elements from hydrogen to curium.

🌱 As universal MLIPs move into practical use, benchmarks have to cover long-time behavior and agreement with experiment, not regression accuracy alone. MLIP Arena is a pioneering effort in that direction, and it aligns with where we are taking PFP.

Read the overview: https://t.co/YZ6nJIbBQI

#Matlantis #MLIP #ComputationalChemistry

[Matlantis™ @matlantis_en]

📢 Matlantis will give two presentations at the 75th Symposium on Macromolecules, September 2-4 at Tokyo University of Science, Katsushika Campus.

Both talks cover what a universal machine learning interatomic potential brings to polymer research: generating polymer-metal interfacial adhesion data at a scale DFT cannot reach and interpreting the mechanism behind it, and how far the accuracy holds once autonomous exploration takes over.

🗓️ Both talks are on Thursday, September 3
- 10:15-10:30, Session C Polymer Function, Bon Cho: Extensive Data Generation and Mechanism Interpretation of Polymer-Metal Interfacial Adhesion Using a Universal Machine Learning Interatomic Potential
- 16:15-16:30, Session B Polymer Structure and Polymer Physics, Akihiro Nagoya: Evaluating Accuracy and Autonomous Discovery using Machine Learning Interatomic Potential

📍 If you will be at the symposium, come find us after either session.

Event page: https://t.co/lWja8EZuDg

#Matlantis #PolymerScience #MLIP

[Matlantis™ @matlantis_en]

🎓 Two of our team spoke at Doctor's Café #37 & AI Salon #14, hosted on July 22 by the FLOuRISH Institute at Tokyo University of Agriculture and Technology.

Fumio Horino, VP of Marketing, covered how AI-driven simulation is changing materials discovery in batteries, catalysts, and semiconductors. Yusuke Asano, VP of Tech Solution, spoke on why crossing fields matters, using his own career as the example: semiconductor materials R&D, a visiting position at UT Austin, materials informatics at ENEOS, and now Matlantis.

💡 In the panel, Asano divided a researcher's work into three abilities: finding problems, solving them, and managing them. His argument was that AI now supports the solving, which leaves the finding as the core of the job.

Read the full report: https://t.co/uh474xO5Ks

#Matlantis #AIforScience