Our comprehensive i̶n̶-̶h̶o̶u̶s̶e̶ defect modelling python package 𝙙𝙤𝙥𝙚𝙙 is now fully live!⚛️. 𝙙𝙤𝙥𝙚𝙙 implements the defect.simulation workflow in an efficient and user-friendly, yet powerful and fully-flexible, manner. See 🧵 for features (1/n).

I think this shows exciting potential for MLFFs in defect modelling, but with caveats. they fail dramatically for non-fully-ionised charge states where localisation matters! .They work here due to the enormous configuration space but with relatively simple underlying energetics.

I find that foundation models (MACE, NequIP, Allegro -- stayed tuned for the latter!) successfully predict split vacancy formation in most cases. This allows an efficient tiered screening; scanning all compounds in the ICSD & Materials Project database for split cation vacancies

The search space still remains too large for DFT however, if we want to go beyond this subset of materials. Due to the relatively simple underlying energetics (primarily electrostatics and strain), this problem is well-suited to universal/foundation ML potentials. .

This allows screening for split cation vacancies in a subset of stable metal oxide compounds (, finding these defect configurations to be far more prevalent than previously known

But we also can't brute-force enumerate potential split vacancy configurations, as the search space is enormous (>1000s of candidate geometries per defect). That said, I find that electrostatic models can greatly reduce this space (as electrostatics dominate energetics here)

However, they are very challenging to identify with current defect structure-searching methods (e.g. ShakeNBreak) due to their 'non-local' nature, as most of these methods employ some form of 'local' structure searching techniques.

Vacancy defects can sometimes transform to split-vacancies, with dramatic changes in energy & behaviour, e.g. in Ga₂O₃ as discovered by Joel Varley. They have only been witnessed in a handful of cases – are they inherently rare or have we just not had the tools to find them?

Article link:. doped (with electrostatics & split vacancies code in split_vacancies branch now): ShakeNBreak:

Machine learning can be powerful for understanding defects, but currently sufficient only in select cases. MLIPs (& geometric/electrostatic tools in doped) allow screening for challenging 'non-local' defect reconstructions (split vacancies) in all ICSD/MP solids, w/caveats 🔗

Collaboration with @scanlond81 @unibirmingham @UCLChemistry @ImpMaterials @lonepair @la_UPC.

Starting with a visit to London in 2023, Cibrán began a deep dive on defects in pnictogen chalcohalides (BiChX), finding the chalcogen vacancy to dominate recombination (similar to Sb2Se3!). He shows that selective anion substitutions can mitigate their effect!.(🔗 below)

RT @Materials_Intel: Last month, we released a major update to the NequIP framework that fully leverages PyTorch 2.0 compilation for MLIPs.….

RT @kanta_v: Thanks to @Kavanagh_Sean_’s huge help, this research came to life👻.

@kanta_v Other computational packages used include ShakeNBreak, CarrierCapture.jl, nonrad, sumo and pymatgen.

@kanta_v Using the recently-added site competition handling for concentration analyses in doped.

Check out @kanta_v's work investigating the mechanism by which Al doping suppresses non-radiative recombination in SrTiO₃, to allow high quantum efficiencies and champion photocatalytic peformance! 📈. Spoiler; it's all in the orbitals 🛰️🤙

RT @kanta_v: 欠陥耐性制御に関する論文が@J_A_C_S に掲載されました。高活性光触媒SrTiO3:Alにおける、Alドープの役割を第一原理計算から調べ、Alはキャリアのトラップ準位の原因となる軌道相互作用を非活性化し、欠陥耐性を高めることを明らかにしました。@K….

RT @jrib_: Matbench Discovery is out in Nature Machine Intelligence @. Paper: Leaderboard: .

RT @rdrons: Proofs are getting worse and worse. What the largest for-profit scientific publisher (with 8 letters) has been doing for years,….