GHOST Meeting Notes

2025-01-03

• @mmoderwell: working on temperature ramping simulations. Taking inspiration from the GNoME paper, attempting to use ASE and MLIP models (and DFT) to simulate how a material will change as it's heated up to room temperature.

  Material screening pipeline

  Superconductivity typically emerges from strong interactions between electrons and vibrations in the crystal lattice (phonons). These interactions can lead to electron pairing, enabling resistance-fre

  @mmoderwell

  #superconductorsDec 27

  • Try to compare MLIP vs DFT calculations if possible

  • Visualize how the lattice changes with temperature

  • See if we can estimate stability

    • Take materials we know to decompose at certain temps and replicate that in silico

• @will: working on replicating out of sample MLIP performance from the GNoME paper using NequIP. 

  • Derive ionic diffusion coefficients for super-ionic conductors.

  • If the results match the GNoME findings, attempt to model a known a superconductor or HEA and compare the MLIP simulation findings with AIMD or other physics based molecular dynamics models.

2025-01-10

• @mmoderwell: working to improve simulations by using some of the better MLIP models (ranked here), like MatterSim from Microsoft or ORB. Read the MatterSim paper and they talk about direct property prediction from their graph model, see if we can better understand how that's done.

  • Looking into fine-tuning MLIP models. Most are trained on GGA-PBE (and sometimes with Hubbard U correction) but there are other levels of theory that may be better suited to superconductivity research. The MatterSim paper demonstrates fine-tuning on other methods with relatively small amounts of data.

  • ORB paper and tools talks about training with D3 corrections, and the benefit of training the model at the level of theory you want instead of adding corrections as an independent step, especially considering some of the computational requirements for said corrections.

2025-01-17

• @mmoderwell working on different approaches to studying superconductivity and predicting Tc in silico. Direct Tc prediction isn't working well, for good reason. On approach to test will be running MD temp ramping simulations and training a classifier to predict if a materials is in the superconducting state, or not. This way we at least don't have to make assumptions about structural changes as a temperature nears it's critical temperature. Can still use MLIP to speed this up. We'd better learn the structural/electronic conditions which create the superconducting state this way too which we could further study if the model works. Not as good for high-throughput screening because of the need for simulating materials heating to room temp or up until greater than Tc.

  • Looking into studying optical and topological properties of superconductors around their critical temperature. Developing some theories about the mechanisms underlying the different pairing mechanisms we see in different classes of superconductor, potentially getting closer to a unified theory.