Multi-Agent Collaboration for Automated Design Exploration on High Performance Computing Systems
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: Next we'll be talking about the paper "Multi-Agent Collaboration for Automated Design Exploration on High Performance Computing Systems".
Jane: The paper was written by Harshitha Menon, Charles F. Jekel, Kevin Korner, Brian Gunnarson, Nathan K. Brown et al. from Lawrence Livermore National Laboratory and Sandia National Laboratories and Los Alamos National Laboratory.
Tom: Stay tuned as we take you through the paper and discuss its implications.
Title: Tom: We've got a massive one to kick things off today, Jane! The paper is called Multi-Agent Collaboration for Automated Design Exploration on High Performance Computing Systems.
Jane: It sounds incredibly technical, doesn't it?
Tom: It really does, but the core idea is actually pretty beautiful once you strip away the jargon.
Jane: Right, basically it's about using a team of AI agents to handle the heavy lifting of scientific design on those massive supercomputers.
Tom: And we've got authors from some of the biggest names in science, like Lawrence Livermore and Sandia National Laboratories.
Jane: That tells me this isn't just a theoretical exercise; they are working with real-world physics problems at the highest level.
Lu: Can you imagine the scale of what they could do? We aren't just talking about designing a simple part, but potentially optimizing entire fusion reactor configurations!
Meng: I wonder how much friction there is when you try to plug an LLM into those old-school national lab supercomputers.
Jane: That's exactly what the paper addresses, Meng, by creating a bridge between the AI and the hardware.
Meng: If they can actually make that connection reliable, it could change how we deploy engineering workflows entirely.
Lalam: It feels like we are moving toward a culture where scientists act more like conductors of an orchestra rather than manual laborers.
Tom: That's a great way to put it, Lalam!
Jane: Instead of spending weeks setting up simulations, the agents handle the orchestration.
Tom: We should probably look at how these agents actually work together in the next segment.
Summary: Jane: So, we've established that this paper, Multi-Agent Collaboration for Automated Design Exploration on High Performance Computing Systems, is about automation.
Tom: Let's break down this MADA framework they built, because it uses three very specific types of agents.
Jane: There's the Job Management Agent, which I like to think of as the project manager for the supercomputer.
Tom: Right, it talks to a scheduler called Flux to make sure all those massive simulation jobs actually run.
Jane: Then you have the Geometry Agent, which is like a digital architect using tools like Cubit or PMesh to build the shapes.
Tom: And finally, there's the Inverse Design Agent, which acts as the brain that looks at the results and says, "Hey, let's try this instead."
Meng: How do these agents actually talk to those specific geometry and simulation tools without breaking everything?
Jane: They use something called the Model Context Protocol, or MCP.
Meng: So it's a standardized way for the AI to call functions in a specialized piece of software?
Tom: Exactly, it lets the agent "plug in" to the tool just like you'd plug a mouse into a computer.
Lu: This is so much more powerful than just asking an LLM to write code; it's giving the AI actual hands to touch the simulation environment!
Lalam: It creates this loop where the AI isn't just guessing, but is actually learning from the physical reality of the simulation.
Tom: And that loop is being used to solve a really tough problem called Richtmyer–Meshkov Instability.
Jane: We'll talk about why that specific physics problem is such a big deal in just a moment.
Improvements: Tom: We are getting into the meat of the results now, and the numbers are actually quite impressive.
Jane: They used this MADA system to try and suppress RMI, which is a type of instability that can mess up fusion ignition.
Tom: And they managed to reduce that instability by about ten percent compared to their starting point!
Jane: That might not sound huge to some people, but in the world of fusion energy, a ten percent improvement is absolutely massive.
Meng: I'm curious about the efficiency part—did it take way more computing power to get that result?
Tom: Actually, they showed that when they used a machine learning surrogate model, the agent could find optimal designs incredibly fast.
Jane: They even compared the agent's performance to traditional mathematical optimization methods like L-BFGS-B.
Tom: And the agent actually found better results in fewer tries than some of those standard methods!
Meng: That makes a lot of sense if the agent can use reasoning to skip over bad parts of the design space.
Lu: It's not just about speed, though; it's about the fact that the agent explains *why* it's making a choice.
Jane: Yes, they showed these "reasoning traces" where you can see the agent thinking through its steps.
Lalam: That transparency is what will build trust in scientific AI, allowing humans to verify the logic behind a discovery.
Tom: It really turns the whole process of design exploration on its head.
Conclusion: Jane: We've covered a lot of ground today with Multi-Agent Collaboration for Automated Design Exploration on High Performance Computing Systems.
Tom: From the specialized agents like JMA and GA to that ten percent improvement in suppressing RMI, this is a huge step forward.
Jane: It's clear that combining LLM reasoning with heavy-duty HPC tools is the future of rapid scientific discovery.
Lu: I can't wait to see these agents designing next-generation materials or even complex spacecraft components!
Meng: If we can make this robust enough for production, it's going to save researchers an incredible amount of time.
Lalam: It truly represents a new era where human intuition and machine precision work in perfect harmony.
Tom: Well, that's all the time we have for this one!
Jane: Thanks for listening, everyone!
Tom: We'll catch you on the next paper!
Lawrence Livermore National Laboratory · Sandia National Laboratories · Los Alamos National Laboratory
cs.AI
Submitted: 2026-03-12
Updated: 2026-09-15
Code: https://github.com/CEED/Laghos
Project page: https://docs.trychroma.com
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: This paper presents MADA (Multi-Agent Design Assistant), an LLM-powered multi-agent framework designed to automate complex scientific design workflows on high-performance computing (HPC) systems.
Key concepts
- MADA Framework
- A system using three specialized AI agents to automate scientific design on supercomputers. It includes a Job Management Agent for scheduling via Flux, a Geometry Agent for building shapes using tools like Cubit or PMesh, and an Inverse Design Agent that analyzes results to suggest new design improvements.
- Model Context Protocol (MCP)
- A standardized method that allows AI agents to connect with specialized software. It acts as a bridge between the AI and simulation tools, enabling the agent to 'plug in' and call specific functions within the software environment, much like plugging a mouse into a computer.
- Richtmyer–Meshkov Instability (RMI)
- A physical instability that can interfere with fusion ignition. The MADA framework was tested by attempting to suppress this instability, ultimately achieving a ten percent improvement compared to initial designs, which is considered a massive advancement in the context of fusion energy research.
Terminology
Summary
This paper presents MADA (Multi-Agent Design Assistant), an LLM-powered multi-agent framework designed to automate complex scientific design workflows on high-performance computing (HPC) systems. As scientific discovery increasingly depends on exploring vast design spaces that are too large and complex for manual exploration,
MADA addresses the need for closed-loop workflows that tightly integrate simulation, analysis, and design exploration
to accelerate the pace of discovery.
The MADA Architecture
MADA modularizes the scientific design workflow into specialized agents that act as reasoning engines that guide collaboration by analyzing results, extracting insights, and adapting workflows as the context evolves.
The framework is built on top of the AutoGen platform and utilizes the Model Context Protocol (MCP) to allow agents to directly call domain-specific tools
through a standardized interface. This integration enables a seamless transition from high-level user objectives to executable actions across diverse computational environments.
The framework is organized around several core components:
-
Specialized Agents: Responsible for distinct stages of the scientific workflow.
-
Planning: Decomposes high-level scientific goals into subtasks.
-
Coordination: Enables agents to communicate and share results through well-defined interfaces.
-
Tool Integration: Uses MCP to expose mesh generators, simulation codes, and HPC schedulers.
-
Memory: Ensures results and context persist for
continuous improvement over iterative design cycles.
Specialized Agent Capabilities
The framework decomposes the design loop into three primary agents, each managing a specific domain of the scientific process. The Job Management Agent (JMA) interfaces with simulation codes and schedules job ensembles on HPC systems,
specifically using Flux for resource management and Laghos for hydrodynamics simulations. The Geometry Agent (GA) automates mesh generation and validation
by translating plain-text CAD descriptions into commands for tools like Cubit or PMesh, utilizing a Retrieval-Augmented Generation (RAG) pipeline to access documentation.
The third component, the Inverse Design Agent (IDA), is responsible for optimization and design-space exploration.
The IDA performs the following tasks:
-
Calculates Quantities of Interest (QoI) from simulation outputs to guide decisions.
-
Ranks candidate designs based on their performance.
-
Proposes new candidate designs, either through broad exploration or
targeted sampling near promising regions.
-
Interfaces with machine learning surrogate models for
rapid design exploration
when full simulations are too computationally expensive.
Validation through RMI Suppression
The researchers validated MADA by focusing on Richtmyer–Meshkov Instability (RMI) suppression, a critical challenge in Inertial Confinement Fusion. The evaluation was conducted in two settings: running full hydrodynamics simulations on the Tuolumne HPC system and using a pre-trained machine learning surrogate model. In the full simulation setting, MADA successfully executed iterative design refinement, achieving a 10% improvement in the Quantity of Interest (reducing QoI from 4.1 to 3.7) within approximately 40 minutes.
In the surrogate-based evaluation for high-velocity impact, the results demonstrated:
-
The agent reached the global optimum for minimizing RMI in
less than 40 evaluations,
matching gradient-based optimization. -
The IDA provided
interpretable reasoning
regarding why certain configurations performed well, such as identifying thatsign variation is critical
for driving RMI growth. -
The system successfully identified that the best designs often sit on the boundary of the admissible range.
Improvements for AI systems
1. Integration of Bayesian Optimization (BO) within the Inverse Design Agent (IDA)
-
Improvement: Replace the current sampling-based approach (Latin Hypercube Sampling) with a Gaussian Process-based Bayesian Optimization acquisition function integrated directly into the LLM's reasoning loop.
-
Capability: The improved AI system can navigate high-dimensional, non-convex design spaces with significantly fewer high-fidelity HPC simulation calls by mathematically balancing exploration of unknown regions and exploitation of known optimal regions.
2. Implementation of a Formal Constraint Verification Agent
-
Improvement: Add a specialized
Verifier Agent
that uses symbolic mathematics or formal logic to intercept the IDA’s proposed design parameters before they are passed to the JMA. -
Capability: The system can perform
Physical Feasibility Pre-screening,
ensuring that all proposed designs strictly adhere to complex physical constraints (e.g., non-negative internal energy, boundary stability, or material limits), thereby eliminating wasted HPC compute cycles on physically impossible configurations.
3. Automated MCP Wrapper Generation via Code Introspection
-
Improvement: Develop an LLM-driven
Interface Synthesis Engine
that can ingest raw source code (C++, Python, Lua) and automatically generate the JSON-RPC 2.0 compliant Model Context Protocol (MCP) servers required for tool integration. -
Capability: This enables
Zero-Shot Tool Onboarding,
allowing the AI system to integrate entirely new, legacy, or proprietary simulation codes and hardware schedulers into a multi-agent workflow without manual developer intervention.
4. Uncertainty-Quantification (UQ)-Driven Adaptive Fidelity Orchestration
-
Improvement: Implement an active learning controller that monitors the predictive uncertainty of the surrogate models used by the IDA.
-
Capability: The system can execute
Dynamic Fidelity Switching,
where it automatically decides whether to perform a rapid, low-cost surrogate evaluation or a high-cost, high-fidelity Laghos HPC simulation based on whether the current design's predictive uncertainty exceeds a predefined threshold.
5. Topological Gradient-Aware Geometry Refinement
-
Improvement: Enhance the Geometry Agent (GA) by augmenting its topological graph representation with a differentiable geometric engine that allows for local sensitivity analysis of mesh quality metrics.
-
Capability: The improved system can perform
Gradient-Guided Mesh Optimization,
enabling it to sense how infinitesimal changes in CAD boundary parameters will affect downstream simulation stability, leading to much faster convergence in generating high-quality, simulation-ready meshes.
Abstract
Today's scientific challenges, from climate modeling to Inertial Confinement Fusion design to novel material design, require exploring huge design spaces. In order to enable high-impact scientific discovery, we need to scale up our ability to test hypotheses, generate results, and learn from them rapidly. We present MADA (Multi-Agent Design Assistant), a Large Language Model (LLM) powered multi-agent framework that coordinates specialized agents for complex design workflows. A Job Management Agent (JMA) launches and manages ensemble simulations on HPC systems, a Geometry Agent (GA) generates meshes, and an Inverse Design Agent (IDA) proposes new designs informed by simulation outcomes. While general purpose, we focus development and validation on Richtmyer--Meshkov Instability (RMI) suppression, a critical challenge in Inertial Confinement Fusion. We evaluate on two complementary settings: running a hydrodynamics simulations on HPC systems, and using a pre-trained machine learning surrogate for rapid design exploration. Our results demonstrate that the MADA system successfully executes iterative design refinement, automatically improving designs toward optimal RMI suppression with minimal manual intervention. Our framework reduces cumbersome manual workflow setup, and enables automated design exploration at scale. More broadly, it demonstrates a reusable pattern for coupling reasoning, simulation, specialized tools, and coordinated workflows to accelerate scientific discovery.
Sources
- GPT-4 Technical Report
- ChemCrow: Augmenting large-language models with chemistry tools
- Galactica: A Large Language Model for Science
- LLaMA: Open and Efficient Foundation Language Models
- SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models
- ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs
- Voyager: An Open-Ended Embodied Agent with Large Language Models
- ChatDev: Communicative Agents for Software Development
- Building Cooperative Embodied Agents Modularly with Large Language Models
- Differentiable Lagrangian Shock Hydrodynamics with Application to Stable Shock Acceleration of Density Interfaces
- Leveraging Vision-Language Models for Manufacturing Feature Recognition in CAD Designs
- Generating CAD Code with Vision-Language Models for 3D Designs
- Are LLMs Ready for Real-World Materials Discovery?
- Multi-Agent Design Assistant for the Simulation of Inertial Fusion Energy
- Barbarians at the Gate: How AI is Upending Systems Research
- URSA: The Universal Research and Scientific Agent
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems
- VSAL: A Vision Solver with Adaptive Layouts for Graph Property Detection