Physics of Soft Sphere Packing + Human Engineers in the Age of AI

In this meetup, we'll talk about some hot topics and some cool topics. We will also have 🍕 pizza!

We welcome Praharsh Suryadevara, a Physics PhD student talking about soft sphere packing, and Dr. Chris Rackauckas, talking about the role of human engineers in an increasingly AI-centric world.

Time
6-8pm, February 18.

Location
Star Room (32-D463), MIT Stata Center. Message on the `#boston-local` channel on the Julia Slack if you are lost.

First Talk:
Structure of basins of attraction of soft sphere packings
Speaker: Praharsh Suryadevara is a PhD student in Physics studying soft matter at NYU, and recently defended his thesis supervised by Stefano Martiniani. He works on the energy landscapes of disordered systems, particularly those of jammed systems and glasses.

The energy landscape picture is a central tool to study many-body systems. In particular, the energy landscapes of glass-forming liquids, jammed packings, constraint satisfaction problems, or neural networks contain a plethora of minima corresponding to competing states. Due to their complexity, these landscapes resist analytical treatment and must be studied numerically. In the first part of my talk, using jammed soft spheres, a paradigmatic example I will show how analyses of steepest-descent basins of attraction in the energy landscapes of liquids and glasses relied on flawed numerical methods, and how DifferentialEquations.jl's benchmarking capabilities enabled identifying these issues. After addressing these methodological issues, in the second part I will show their impact on two notions of global stability for attractors based on their basins of attraction in high dimensions, one based on the fractality and one based on volumes, with consequences for previous claims in the field.

Second Talk
The Bitter Lesson for the Bitter Lesson: The Role of Human Engineers in the Age of AI
Speaker: Dr. Chris Rackauckas, of DiffEq, SciML and Dyad fame

With the ever increasing role of AI it is pertinent to ask the question: what will be the role of engineers and domain experts? In this talk we outline two approaches that incorporate the expertise of engineers into the latest techniques of AI and machine learning. First we showcase scientific machine learning, in particular universal differential equations, where machine learning architectures are mixed with traditional simulation techniques in order to discover higher order physical corrections and generate hypotheses for previously unknown governing laws. Then we dive into growing techniques for agentic AI in modeling and simulation, highlighting the recent empirical results around the tools and techniques which lead to improved accuracy of code generation in the context of nonlinear controls synthesis and analysis. Building on these results, we showcase the new Dyad platform for agentic AI which combines a new statically-analyzable acausal equation-based DAE modeling system with built-in scientific machine learning capabilities and demonstrate the real-world applications this is seeing in industrial applications from aerospace and automotive all the way to chemical process modeling. This talk will span the details from low level mathematical theorems to high level software demonstrations in applications, highlighting the elements of the new stack which have successfully translated into practice but also the areas which need further academic work to complete the transition.