@all

Core Idea: Train a GNN from scratch to predict MAE using CHGNet-derived features: Node features: CHGNet latent embeddings (structural context) + CHGNet magmom predictions (explicit magnetic state)

This asset shows two plots for a CuNi crystal after a 10 picosecond melt equilibration at 1800 K. The left plot is the total radial distribution function (RDF) versus distance, with a strong first peak near 2 Å and several smaller peaks up to about 8–9 Å, suggesting some remaining order from the original lattice. The right plot shows the coordination number (CN) as a function of distance, which increases gradually and reaches around 350 by 10 Å. The note indicates that even at about 9 Å away, there is still a signal of another atom, meaning remnants of the supercell lattice persist in the melted state.

I'm excited to share a new page I've been building out this week. You may have already seen it, as it's the first page you be redirected to after sign in.

is a post about running molecular dynamics simulations to study how a Cu-Zr alloy forms a metallic glass. The author uses a 64% Cu and 36% Zr composition, an (10,10,10) supercell, and the orb-v3-direct-20-omat calculator to push speed and scale. The workflow includes equilibrating a melted alloy at high temperature, then rapid quenching from 2000 K to 300 K at various rates to compare glass formation versus crystallization. The write-up explains key concepts like what glass is in atomic terms, the difference between crystalline order and amorphous structure, and how RDF and coordination numbers help analyze results. It also notes the challenges of achieving crystallization in MD due to time scales and suggests exploring different cooling rates and compositions in future runs. The post includes example data and 3D visualization references to support the findings.

explains how to pick from eight Orb-v3 models that balance accuracy, speed, and memory for atomistic simulations. The post breaks down model names (orb-v3-X-Y-Z), where X is how forces are computed, Y is neighbor limits, and Z is the training dataset (omat or mpa). It compares conservative vs direct force calculations, unlimited vs limited neighbors, and AIMD-based -omat versus MPTraj/Alexandria-based -mpa models. Readers gain practical guidance for phonon calculations, geometry optimization, and molecular dynamics, including which models excel at energy conservation, speed, or large-scale simulations. The piece also covers workflow tips, performance at scale, and licensing (Apache 2.0). Use this guide to choose the right Orb-v3 model for your system size and research goals.

The Pareto frontier for a range of universal Machine Learning Interatomic Potentials. The 𝐾𝑆𝑅𝑀𝐸 metric assesses a model’s ability to predict thermal conductivity via the Wigner formulation of heat transport and requires accurate geometry optimizations as well as second and third order derivatives of the PES (computed via finite differences). The y-axis measure a model’s forward passes per second on a dense periodic system of 1000 atoms, disregarding graph construction time, measured on a NVIDIA H200. Point sizes represent max GPU memory usage. Y-axis jitter (+/- 5 steps/second) has been applied to allow visualization of overlapping points. Model families include a range of specific models with broadly the same architecture, but may be different sizes or trained on different datasets.

Meso-scale all-atom simulations with Orb-v3 open a new frontier in materials science and chemistry. This post discusses using ASE for molecular dynamics on GPUs (A100, H100, H200) via Modal, enabling larger, more affordable simulations than traditional DFT. It highlights metallic glass formation, crystallization, annealing, and emergent phenomena that arise from thousands of atoms. The focus is on a breakthrough in non-conservative architectures that balance memory use and speed, making complex systems feasible to study.

The authors introduce Orb-v3, the next generation of the Orb family of universal interatomic potentials. Models in this family expand the performance-speed-memory Pareto frontier, offering near SoTA performance across a range of evaluations with a ≥ 10× reduction in latency and ≥ 8× reduction in memory. Their experiments systematically traverse this frontier, charting the trade-off induced by roto-equivariance, conservatism and graph sparsity. Contrary to recent literature, they find that non-equivariant, non-conservative architectures can accurately model physical properties, including those which require higher-order derivatives of the potential energy surface.

Paper by Rich Sutton

Overview of the current work and future enhancements for magnetic hysteresis simulations

A post about trying to use HamGNN with TB2J to forecast magnetocrystalline anisotropy energy, only to find the pre-trained model lacks the needed physics. The main gap is the absence of spin-polarization in H0, making the model better suited for SOC in non-magnetic materials, not for magnetic predictions. Potential outputs still relevant to SOC include band structure corrections, topological invariants, spin textures in k-space, orbital angular momentum, spin Hall conductivity, g-factors, effective masses, and optical properties. The use case focuses on non-magnetic materials and topological insulators without magnetism. Next steps involve exploring new Hamiltonian models like DeepH-pack and MACE-H, noting they lack pre-trained models. The plan is to gather consistent data, ensure SOC and spin-polarization, and align data sources from the same DFT software. Links: https://github.com/mzjb/DeepH-pack, https://github.com/maurergroup/MACE-H.

Phase diagram of Mn8Al8C; e_above_hull: 0.315834 eV/atom; predicted_stable: False

Cell + Ionic relaxation with Orb v3; 0.03 eV/Å threshold; final energy = -111.2538 eV; energy change = -38.8208 eV; symmetry: P4/m → P1

Crystal structure generated by GEPA optimization (iteration 2)

Phase diagram of Mn8Al8C; e_above_hull: 1.072304 eV/atom; predicted_stable: False

Cell + Ionic relaxation with Orb v3; 0.03 eV/Å threshold; final energy = -85.6817 eV; energy change = -99.9416 eV; symmetry: P4/mmm → P1

Crystal structure generated by GEPA optimization (iteration 1)

Cell + Ionic relaxation with Orb v3; 0.03 eV/Å threshold; final energy = -7.2183 eV; energy change = 2.4552 eV; symmetry: P4/mmm → P4/mmm

Crystal structure generated by GEPA optimization (iteration 12)

Phase diagram of Fe4CoSiB2; e_above_hull: 0.173824 eV/atom; predicted_stable: False

Cell + Ionic relaxation with Orb v3; 0.03 eV/Å threshold; final energy = -245.7428 eV; energy change = -455.6124 eV; symmetry: I4/mcm → P1

Crystal structure generated by GEPA optimization (iteration 11)

Cell + Ionic relaxation with Orb v3; 0.03 eV/Å threshold; final energy = -24.1489 eV; energy change = -0.9037 eV; symmetry: P4/mmm → P1

Crystal structure generated by GEPA optimization (iteration 10)

Phase diagram of Fe2B; e_above_hull: 0.000062 eV/atom; predicted_stable: True

Cell + Ionic relaxation with Orb v3; 0.03 eV/Å threshold; final energy = -98.1390 eV; energy change = -9.9987 eV; symmetry: I4/mcm → I4/mcm

Crystal structure generated by GEPA optimization (iteration 9)

Dataset BTC-USD downloaded from yfinance: 2020-01-01 to present

Dataset BTC-USD downloaded from yfinance: 2020-01-01 to present

Dataset BTC-USD downloaded from yfinance: 2020-01-01 to present

Once again we're at a stopping point because of our inability to effectively predict MAE. Our AI discovery agents have discovered materials that have all the properties we can currently predict. This

A clear look at using LLMs for materials discovery. This post summarizes how Victor from Lila Sciences and the author compare their AI workflows, focusing on mutations in crystal structures, machine learning force fields (MLFF), and machine learning property predictors (MLPP). It explains why data limits push researchers toward direct reasoning with AI, the role of modification history and self-reflection, and how prompts influence performance. Key takeaways include constraints in rare-earth-free magnet design, the balance between control and relaxation, and the impact of self-reflection on speeding up convergence to target properties like band gap and formation energy. The notes also touch on challenges with crystal symmetry, CIF generation, and future work in MLIP/MLPP accuracy. A practical read for anyone exploring automated materials discovery and AI-driven design.

This post explores ideas for finding adjacent crystals in Matra-Genoa’s latent space to discover materials with targeted properties. The author describes challenges when mutating crystals, where small input changes can lead to large, different outputs after relaxation. Three approaches are considered: conditioned generation with token hints (fixing some inputs while mutating others), decoding from a modified latent space (using predictors and SHAP to steer latent directions before decoding), and a hybrid approach that combines fixed tokens with latent-space moves. The goal is faster exploration and smarter guidance from an AI research agent and a language model, reducing the cost of property evaluation. The notes also touch on fine-tuning and property-focused training to improve material design workflows. Keywords: adjacent crystals, latent space, Matra-Genoa, crystal generation, materials AI, property optimization.

Check out the paper here. It's a short read. I recommend checking it out. Although not very technical (just machine learning concepts that have been explored elsewhere), the creativity and simplicity

AI-discovered magnetic material: Mn2CrFe4Co4N (performance score: 0.740) | Space group: 1 (resolved from structure) | Key properties: Tc: 612K, Ms: 0.14T, Cost: $13/kg, E_hull: 0.235eV/atom, Dynamically stable | Discovered in 20 AI iterations | - The combination of Mn, Cr, Fe, Co, and N in this stoichiometry yields a high Curie temperature and magnetic density. - The material is dynamically stable, which supports its structural integrity. - The energy above hull suggests that the material is metastable or unstable thermodynamically. - Cost is low, indicating practical feasibility from an economic standpoint.

AI-discovered magnetic material: Fe11CoSiGeAsP (performance score: 0.597) | Space group: 8 (resolved from structure) | Key properties: Tc: 687K, Ms: 0.13T, Cost: $82/kg, E_hull: 0.161eV/atom, Dynamically stable | Discovered in 10 AI iterations | - Strong ferromagnetism with high Tc arises naturally from the Fe/Co sublattice; this is retained despite chemical complexity. - Dynamic stability indicates the structure is at least locally stable; the main risk is competition with lower-energy phases (slightly positive e_hull). - The metastability is small enough that slight stoichiometric shifts (e.g., favoring smaller/more covalent anions like P over As, or Si over Ge) or controlled disorder could stabilize the phase thermodynamically. - Magnetic density is adequate but not exceptionally high; maintaining or modestly enhancing it while reducing e_hull should be feasible by delicate tuning of Co content or anion ratios.

AI-discovered magnetic material: Fe4Mn3B4 (performance score: 0.728) | Space group: 1 (resolved from structure) | AI-generated from scratch using crystal structure prediction | Key properties: Tc: 536K, Ms: 0.09T, Cost: $1/kg, E_hull: 0.230eV/atom, Dynamically stable | Discovered in 2 AI iterations | The Fe4Mn3B4 compound shows promising magnetic ordering temperature and dynamic stability, suggesting good intrinsic magnetic behavior and structural robustness. The main challenge is its thermodynamic stability, as indicated by the high energy above hull. The magnetic density is close but slightly below the target, suggesting that minor compositional or structural modifications might improve it. The low cost and atom count within limits make it a practical candidate if stability can be enhanced.

Interactive trajectory explorer with MatterViz

Welcome

Interactive browser visualizations for materials science, by @janosh

Welcome

Relax a crystal structure and create a post

Get a detailed description of a crystal structure

Generate CIF file from crystal structure description

Generate a crystal structure using GGen

Root

Get space groups compatible with a given chemical formula

Random bulk crystal generation with PyXtal and Orb v3

Relax a crystal structure with animation

Create interstitially doped structure

Generate a crystal structure with MatterGen

Generate a crystal structure with Chemeleon

Dataset powering the material cost calculator. Lists element's USD/kg and when the data was last updated and where it came from.

Calculate the estimated raw material cost per kg

Generate crystal structures with magnetic density and HHI score conditioning

Analyze Structure

Calculate Thermochemistry

List Supported Formats

Forecasts for Bitcoin Price with 12-period horizon

Forecasts for Oil Price with 12-period horizon

Forecasts for Gold Price with 12-period horizon

Forecasts for Gold Price with 52-period horizon

Forecasts for Bitcoin Price with 52-period horizon

This dataset has a set of 34,000 ferro/ferrimagnetic materials from Materials Project, their formula, if they include rare earth elements, magnetic moment, volume, magnetic density, a predicted Curie temperature, and cosine distances to some known permanent magnets like NdFeB. Distances are based on a 256 dimension embedding from Orb v2 latent space.

Forecasts for Gold Price with 52-period horizon

Forecasts for Copper Price with 52-period horizon

Observed and forecasted housing market data for April 2025. Includes monthly data and forecasts projecting 12 months into the future.

A collection of 5020 magnetic materials from Materials Project, with estimated magnetic density and predicted Curie temperatures.

Forecasted fred-cbbtcusd-festive-ride from 2025-04-12 to 2025-12-30

Dataset CBBTCUSD downloaded from fred: 2020-01-01 to present

Dataset BTC-USD downloaded from yfinance: 2020-01-01 to present

Dataset BTC-USD downloaded from yfinance: 2020-01-01 to present

Forecasted fred-cbbtcusd-tender-shirley from 2025-04-09 to 2025-12-30

Dataset CBBTCUSD downloaded from fred: 2020-01-01 to present