Ivan Kruglov | Xpanceo: How can AI slash the time and cost of discovering new optical materials?

Can computer simulations truly predict a material's optical properties with experimental accuracy?

The foundation of an AI-driven discovery platform is a high-quality database of material properties. To build this, scientists rely on Density Functional Theory (DFT), but not all DFT methods are created equal. Simple approaches like PBE are inaccurate for optical properties, so more advanced, computationally expensive methods are required to capture the complex physics involved.

To achieve high accuracy, the team employs sophisticated computational techniques like the GW approximation and the Bethe-Salpeter Equation (BSE). These methods are essential for correctly modeling electronic band structures and, crucially, accounting for excitonic effects—the interaction between electrons and holes—which dominate the optical response of many materials. The excellent agreement between GW+BSE calculations and experimental measurements for molybdenum disulfide (MoS2) is shown as proof of the method's predictive power.

Further validation is provided by comparing calculated and experimental Raman spectra for materials like germanium disulfide and palladium selenide. Since Raman spectra are derived from the material's dielectric tensor (an optical property), the near-perfect match between theory and experiment demonstrates the robustness of the computational framework. This high-fidelity data is the critical ingredient for training a reliable and predictive machine learning model.

In this short video, you can learn:
* The hierarchy of DFT methods (PBE, GW, BSE) for calculating optical properties.
* The importance of accounting for excitonic effects for accurate optical predictions.
* How Raman spectroscopy can be used to validate the accuracy of computationally derived optical properties.
📋 **Clip Abstract** This clip delves into the computational methods used to generate the high-fidelity data needed for training machine learning models. It highlights the use of advanced DFT techniques like GW and BSE to achieve experimental-level accuracy in predicting optical properties, validated against both optical and Raman spectra.
🔗 Link in comments 👇

Can a machine learning model trained on just 1,000 materials find a record-breaking new one?

The speaker details the process of building the dataset for their machine learning model. They started with a custom-calculated set of about 100 van der Waals materials, which was insufficient for training a robust model. To expand this, they merged their high-accuracy data with a larger, existing dataset (NoMD), resulting in a combined training set of over 1,000 structures.

With the dataset prepared, a graph neural network (GNN) was trained to predict the birefringence of materials based on their crystal structure. This trained GNN was then deployed to perform a high-throughput screening of massive computational materials databases, including Materials Project and Google's GNoME. The model rapidly evaluated tens of thousands of candidate materials to identify the most promising ones.

The screening process successfully identified molybdenum ditelluride (MoTe2) as a standout candidate with exceptionally high predicted birefringence. This computational prediction was then fast-tracked for experimental validation. The subsequent measurements confirmed that MoTe2 indeed possesses one of the highest birefringence values among all materials studied, demonstrating the power and efficacy of the entire AI-driven discovery pipeline.

In this short video, you can learn:
* The importance of data augmentation by combining custom and public datasets for ML training.
* How graph neural networks (GNNs) are used to screen large material databases like Materials Project.
* The successful discovery and experimental confirmation of molybdenum ditelluride (MoTe2) as a high-birefringence material.
📋 **Clip Abstract** This clip showcases the successful application of an AI-driven workflow to discover a novel high-performance optical material. It describes how a graph neural network, trained on a combined dataset, screened massive databases to identify molybdenum ditelluride, which was then experimentally confirmed to have one of the highest known birefringence values.
🔗 Link in comments 👇