speaking-notes.org

speaking-notes

Up

Description

Version without outline slides.

Frontmatter

Titlepage

• present yourself
  • hi, i am NAME
  • i am a PhD student at GROUP
  • with my PIs NAMES
  • i will talk about TITLE
    • about quantum materials simulation, and why and how we evolve it from physics-based to hybrid physics/AI
• transition
  • I want to tell you about what led me to this position, from all the way back to the beginning

Section Introduction

Slide Opening - The most important problem

• [click slowly 2x, pause]
• Nearly ten years ago, as a freshman, I looked for the most important problem to solve, right now

Slide Anthropocene 1

• [click slowly until “socio-economic trends”, pause]
• I realized that we live in a remarkable age. Industrialization. Health, mobility, information have never been as good and available.

Slide Anthropocene 2

• [click]
• But the ecological outlook for the future has never looked as bleak.
• Civilization has developed in a relatively stable climactic era, our systems are tuned to it. Though we now they existed, we as civilization have never lived through them.

Slide Carbon Duration

• [click until “how long will the change last?”]
• So how long will this period of change last?
• Roughly 2’000 years. Until CO2 has mixed with the ocean and starts turning into calcium carbonate. So as long as modern civilization.
• I heard this saying once. “If cancer is not solved in our lifetime, the world will stay the same, unfortunate as it is. If this problem [point to screen] is not solved in our lifetime, the world will become a different place.” One that our civilization has not yet encountered. But we can avoid that societal risk by decarbonizing.

Slide Energy mix

• [click until “we are 20 percent done”]
• So, how far are we with solving this problem? About 20 percent. Despite all efforts, 80 percent of our energy mix remains fossil fuel based. We are too slow.
• So, the climate challenge is actually an energy challenge.

Slide IT energy use 1

• [click until “Example: The Energy Challenge in IT”]
• Let’s look at one sector as an example. IT.
• Computing power use is growing faster than total power consumption use. That’s a problem.
• [click to “Notable AI Models”]
• It gets worse with the advent of the deep learning era. Those models are too power-hungry.
• [click until “A disruptive technology”]
• You can only do so much to make existing semiconductor technology more energy-efficient. A truly disruptive change would be if you change how computing is done altogether, with quantum materials.
• This is a diverse class of materials that exhibit purely quantum properties at a macroscopic scale and thus allows to use these properties technologically. Devices built with them could scale with respect to size, speed and energy use orders of magnitude better than what we have now.
• Here are two examples from my research group. On the left are skyrmions, little movable knots or swirling patterns of atomic spins.
• On the right MZMs, another kind of quasiparticle that could be used for fault-tolerant quantum computing.
• [click]
• So, we could say that the energy challenge is to a considerable part actually a challenge for new and better materials

Slide materials development duration

• [click]
• The problem is that it is actually very hard to find an develop a new material. Just look at this list.
• This pace of progress is too slow for what we are facing now

Slide 5th paradigm

• [click]
• How do we accelerate materials discovery? With new scientific paradigms.
• With simulation, we can do high-througput screening of thousands of materials. and build databases. Moreover, unlike experimental data, this data is perfectly labeled and only limited by our compute budget, so perfect conditions for training machine learning models to assist or replace simulation.

Slide First Principles WFT and DFT

• [click to “First-principles”]
• Such “First-principles” simulation methods are what we call methods that work without empirical assumptions, based on quantum mechanics alone.
• Density functional theory is such a method. It is a practically useful approximation of the Schrödinger equation and allows to calculate any propetry of a material or molecule based on its structure alone.
• Before LLMs came along, it occupied the largest share of supercomputers worldwide. But even with that, it cannot even scratch the surface of of the traversable materials space, due to its scaling behavior.
• So it is also too slow for what we are facing

Slide Industry research

• [click]
• So it is unsurprising that the big artificial intelligence players are coming into this field.
• Since only last year, Google, Microsoft, Meta and other companies have published large models that disruptively accelerate materials discovery, they say
• So, AI to the rescue
• [click]
• Let me point out a seeming conundrum here. I said computing has a power poroblem, and I propose to solve it with more computing. Then I said that AI makes it worse, and I propose to solve it with more AI. But, if we use this for enabling a distruptive technology like quantum materials, then it changes the game. Potentially for all materials and thus all energy technologies. And hopefully, in time.
• So, happy end?

Slide All-electron DFT, JuDFT

• [click]
• Note quite. These big corporate models are rather coarse-grained.
• Quantum materials require super-high accuracy to meV and lower
• All-electron DFT can do that. There are only a handful, and two of them are developed at Jülich in our group over four decades
• Integrated with in-house multiscale and HT workflow engines (Spirit, AiiDA)
• Below an example of the accuracy of our code compared to other popular DFT codes

Slide The full pipeline

• [click]
• We arrive at our proposed hybrid physics/AI pipeline. I am building models that learn to emulate our JuDFT codes.
• That boils down to engineering or learning representations of the atomic structure that preserve the symmetries of their related properties, which can have any dimensionality.
• We reuse these predictions as better initial guesses for faster convergence of our DFT codes
• And as secondary property predictors to also accelerate multiscale simulation like spin dynamics (like those skyrmions)