Phenomenological Detector Design and Optimization in Vertically-Integrated Differentiable Full Simulations with Agentic-AI

This paper presents the first implementation of AI agents within a vertically-integrated differentiable full simulation framework to autonomously optimize high-energy physics detector parameters, demonstrating their ability to reduce labor and compute while effectively navigating complex design spaces for tasks like dual-readout calorimeter optimization.

The Gist

Imagine you are trying to design the ultimate camera for a super-fast, high-speed race. But this isn't just any camera; it's a camera so complex that it needs to be built, tested, and fine-tuned all at the same time, using a computer simulation that takes days to run for every single tiny change you make.

This paper is about teaching a super-smart AI assistant (an "AI Agent") how to be the chief engineer for this camera, specifically for a particle physics experiment.

Here is the breakdown of what they did, using simple analogies:

1. The Problem: The "Infinite Lego" Puzzle

In particle physics, scientists build massive detectors (like giant cameras) to catch particles moving near the speed of light. Designing one is like trying to build a Lego castle where:

• You have to decide the size of every brick (the geometry).
• You have to decide how fast the camera shutter clicks (the digitization).
• You have to decide how the computer software groups the pictures together (the reconstruction).

Usually, humans have to guess which combination works best. If you change the brick size, you might have to re-simulate the whole thing. If you change the shutter speed, you have to re-simulate again. It's slow, expensive, and requires a lot of human guesswork.

2. The Solution: The "AI Architect"

The authors created a special system where an AI Agent (powered by a large language model, like a very advanced chatbot) takes the wheel.

Think of this AI Agent not just as a calculator, but as a project manager who can:

• Read the blueprints.
• Change the settings on the simulation.
• Run the test.
• Look at the results.
• Decide what to change next.

They used a "Bilevel Optimization" framework. Imagine a Russian nesting doll:

• The Outer Doll (Geometry): The AI changes the physical shape of the detector (how big the crystals are, how they are arranged).
• The Inner Doll (Software): Once the shape is set, the AI tweaks the software settings (how many bits of data to save, how fast to sample the signal) to get the best result for that specific shape.

The AI does this loop over and over, learning from every attempt.

3. The Experiment: The "Crystal Camera"

They tested this on a specific type of detector called a Dual-Readout Calorimeter.

• The Analogy: Imagine a bucket that catches rain (particles). Usually, you just measure how much water is in the bucket. But this special bucket has two sensors: one that measures the splash (Scintillation) and one that measures the sound of the rain hitting the water (Cherenkov). By comparing the two, you can figure out exactly what kind of rain it was, even if it's a mix of heavy drops and mist.
• The Goal: They wanted to find the perfect size for the "buckets" (crystals) and the perfect settings for the "sensors" to get the clearest picture of the rain.

4. What the AI Discovered

The AI didn't just blindly try random combinations. It started "thinking" and making smart deductions:

• The "Useless Knob" Discovery: The AI noticed that one specific setting (the "projective offset," or how much the crystals are tilted) didn't actually change the quality of the picture much. It realized, "Hey, this knob is a distraction. Let's stop wasting time turning it."
  • Note: The paper admits the AI missed a tiny detail about the scale of the real world (it thought the crystals were huge, when they are actually small), but it still correctly identified that the tilt didn't matter for the test it was running.
• Breaking the Problem Down: Instead of trying to solve a giant 11-dimensional puzzle all at once (which is like trying to solve a Rubik's cube while juggling), the AI realized it could split the problem.
  • Step 1: First, just figure out the best size for the crystals to get a clear signal.
  • Step 2: Once the size is fixed, figure out the best camera settings (sampling rate, bits) to save money and power.
• The Result: The AI successfully navigated the complex space, finding a design that was just as good as (or better than) what a human team would have found, but it did so by reducing the work and saving computer time.

5. Why This Matters

This is the first time an AI has been used to design a particle detector from scratch, rather than just analyzing data from one that already exists.

• The "Human-in-the-Loop": The AI isn't a magic wand that knows physics better than a professor. It still needs a human to give it the initial plan and check if its ideas make sense in the real world.
• The Future: This proves that we can use AI agents to handle the boring, repetitive, and complex math of designing future experiments. It frees up human scientists to focus on the big, creative ideas while the AI handles the heavy lifting of optimization.

In a nutshell: The authors taught an AI to be a master architect for a particle detector. The AI learned to ignore useless settings, break a massive problem into small, manageable steps, and find the best design faster than a human team could do alone. It's a major step toward using AI to build the next generation of scientific discovery machines.

Technical Summary

1. Problem Statement

High-energy physics (HEP) experiments require complex, multi-stage detector designs that integrate geometry, front-end electronics, and reconstruction algorithms. Traditionally, optimizing these systems involves:

• High-dimensional parameter spaces: Combining detector geometry (e.g., crystal size, offset) with digitization settings (e.g., ADC bits, sampling rate) and reconstruction parameters.
• Computational cost: Full simulations (e.g., Geant4) are computationally expensive, making exhaustive grid searches impractical.
• Human dependency: Effective optimization often relies on human experts to prune the parameter space based on domain knowledge, as generic optimization tools struggle with complex correlations and sparse data landscapes.
• Lack of autonomy: While AI/ML is used in data analysis, its application to the design cycle itself—specifically using autonomous agents to navigate the full simulation-to-reconstruction loop—remains unexplored.

The paper addresses the challenge of automating the design and optimization of a detector system using Agentic AI within a vertically integrated differentiable simulation framework, reducing human labor and compute time while maintaining physical rigor.

2. Methodology

The authors propose a novel Bilevel Optimization Framework driven by an Agentic AI system.

A. The Bilevel Optimization Framework

• Outer Loop (Geometry): Iterates over detector geometry parameters using DD4HEP (a toolkit for detector simulation).
• Inner Loop (Reconstruction & Digitization): Evaluates simulation outputs against downstream parameters, including front-end digitization hardware specs and high-level reconstruction algorithms.
• Differentiability: The system supports differentiable full simulations, allowing for the optimization of geometry and algorithm parameters simultaneously.
• Scoring: Results are scored against physics/statistics objectives (e.g., Signal-to-Noise Ratio, fitting errors, power consumption).

B. The Agentic Layer (SciFi)

• Framework: The authors utilize SciFi, a science-focused, open-source agentic AI framework.
• LLM Backbone: The reasoning engine uses Claude Code Opus 4.6 (commercial) for high-level planning, while the execution system uses an open-weight LLM for autonomous task execution.
• Workflow: The agent autonomously:
  1. Proposes test points in the parameter space.
  2. Calls tools to run simulations and analysis.
  3. Interprets returned results.
  4. Summarizes knowledge gained.
  5. Plans subsequent trials based on feedback.
• Design Philosophy: The framework is built with a centralized, human-readable configuration system and modular pipelines to prevent the LLM from regressing into generic, unmaintainable code patterns.

C. Case Study: Dual-Readout Segmented Crystal EM Calorimeter

• Target: A future collider detector (IDEA concept) with a baseline resolution of $3\%/\sqrt{E}$.
• Parameters Optimized:
  • Geometry (3 params): Transverse crystal granularity, projective offset from the Interaction Point (IP), and front-layer crystal lengths.
  • Digitization/Reconstruction (8 params): ADC bits, sampling rate, clustering radius ($R$-size), energy thresholds, and waveform fitting bounds.
  • Total: 11 tunable parameters.
• Objectives: Maximize Energy SNR, minimize detector-level C/S fitting error, and minimize ADC cost (power consumption).

3. Key Contributions

1. First Agentic AI Implementation in HEP Design: This is the first demonstration of an AI agent autonomously executing a vertically integrated workflow that spans from detector geometry definition to high-level reconstruction optimization.
2. Discrete Reasoning Alternative to Differentiable Programs: Instead of relying solely on explicit gradients (which can be difficult to define for complex, non-differentiable simulation steps), the agent uses discrete reasoning and iterative feedback to navigate the parameter space, emulating human expert intuition.
3. Dynamic Parameter Space Reduction: The agent successfully identified "nuisance parameters" (parameters with low sensitivity) and decoupled the optimization problem, reducing a complex 11-dimensional search into manageable sub-problems without explicit human instruction on which parameters to ignore.
4. Vertical Integration: The study bridges the gap between hardware design (geometry), signal processing (digitization), and physics analysis (reconstruction) in a single automated loop.

4. Results

The agent-driven optimization yielded significant insights and performance improvements:

• Identification of Nuisance Parameters:
  • The agent correctly identified that the projective offset was largely insensitive to the SNR metric within the tested range.
  • Note: The agent initially lacked context regarding the physical scale of crystals (0.5–2 cm vs. the tested 1–10 cm), leading to a conclusion that was valid for the simulation range but physically limited. However, it successfully filtered this parameter out of the critical optimization path.
• Decoupling of Optimization Stages:
  • The agent discovered that pre-digitization SNR optimization is independent of the digitization and waveform-fitting parameters.
  • This allowed the 11-dimensional space to be decomposed into two simpler stages:
    1. Stage 1: 2D scan of Crystal Length vs. Width (optimizing SNR).
    2. Stage 2: 2D scan of ADC Bit Depth vs. Sampling Rate (optimizing photon error and cost).
• Adaptive Scanning:
  • The agent adaptively proposed scan ranges, step sizes, and event statistics over rounds A–E, progressively narrowing the optimal region.
  • It identified that fitting parameters were "nuisance" parameters; any reasonable configuration yielded a good fit, meaning they did not require fine-tuning.
• Final Outcome: The system successfully converged on an optimal set of parameters that balanced SNR, fitting error, and power cost, demonstrating that an AI agent can follow an optimization path comparable to a human expert.

5. Significance and Future Outlook

• Efficiency: The approach significantly reduces the labor and compute resources required for detector design by automating the "trial-and-error" loop and intelligently pruning the search space.
• Validation of First Principles: It opens efficient avenues for computationally validating design choices based on first principles rather than heuristic assumptions.
• Current Limitations: The study notes that while agents excel at execution and pattern recognition, they currently lack the ability to make "autonomous leaps of physics-motivated judgment" without human context (e.g., understanding the physical scale of crystals).
• Future Directions: The authors plan to rebase the system entirely on open-weight LLM backbones to ensure full reproducibility and accessibility. They also aim to enhance the agent's ability to incorporate domain-specific context to avoid physical scale errors.

Conclusion: This work defines the current frontier of experimental design in HEP, demonstrating that Agentic AI can effectively manage complex, vertically integrated simulation workflows, transforming detector design from a manual, iterative process into an autonomous, data-driven optimization cycle.