Inaugural Seminar - Assistant Professor Line Jelver

At the inaugural seminar Assistant Professor Line Jelver gave a presentation about Neural Networks for Computational Condensed Matter Physics. She was also joined by two guest speakers from Karlsruhe Institute of Technology (KIT).

• Guest speaker Marjan Krstic (Karlsruhe Institute of Technology/KIT) presents “A scale-bridging framework for (non-)linear optical properties of molecular materials, structures and devices thereof”

• Guest speaker Christof Holzer (Karlsruhe Institute of Technology/KIT) presents “Predicting Light: Quantum Chemistry meets Photonics”

Thanks to guests for joining and celebrating Line Jelver!

Abstract:
Neural networks and, more recently, large language models (LLMs), are becoming powerful tools for accelerating and augmenting computational physics. In this seminar, I will give a gentle introduction to these methods and discuss how they can be integrated into modern condensed matter workflows, with an emphasis on optical response modeling and the automation of first-principles simulations.
First, I will present a deep-learning framework for the forward prediction and inverse design of thin-film multilayer stacks used in nanophotonics applications such as structural color generation, antireflection coatings, and thermally functional surfaces. Designing these stacks is challenging because the mapping from materials and thicknesses to color coordinates is highly nonlinear and interference-driven, while the inverse problem—finding layer configurations that realize a target color—is inherently non-unique. I will show how combining differentiable forward models with population-based optimization enables efficient inverse design within a single, integrated pipeline.
Second, I will describe ongoing work toward data-driven automation of complex first-principles materials modeling workflows. State-of-the-art simulations often require multi-stage procedures, careful convergence testing, and expert parameter choices across several specialized software packages. The goal of this project is to reduce this overhead by learning from prior calculations to recommend simulation strategies, convergence criteria, and parameter settings automatically. As a concrete starting point, I will focus on the entry step of many workflows: Defining the crystal structure. This includes an assessment of how accurately different LLMs can identify a material from user input and return the crystallographic information needed to initialize density functional theory (DFT) calculations