Beyond AI Einsteins: A System for Discovering New Physics

by

Author: Denis Avetisyan

Researchers have developed a novel system that doesn’t just imagine new theories, but actively generates and tests them using a combination of large language models and rigorous simulations.

The FERMIACC model iteratively refines proposed Beyond the Standard Model explanations for experimental anomalies, employing a proposal agent informed by prior results and a critical patching loop before submitting the finalized hypothesis to FeynRules for validation-a process mirroring the scientific method’s cycle of conjecture and verification.
The FERMIACC model iteratively refines proposed Beyond the Standard Model explanations for experimental anomalies, employing a proposal agent informed by prior results and a critical patching loop before submitting the finalized hypothesis to FeynRules for validation-a process mirroring the scientific method’s cycle of conjecture and verification.

FERMIACC leverages scaffolded reasoning and deterministic simulations to autonomously explore the landscape of high-energy physics and validate potential hypotheses.

Despite decades of data from experiments like the Large Hadron Collider, fundamental questions in high energy physics remain unanswered, demanding novel approaches to hypothesis generation and validation. This paper introduces ‘The FERMIACC: Agents for Particle Theory’, a scaffolded reasoning model leveraging large language models and deterministic simulations to autonomously explore the theoretical landscape. FERMIACC moves beyond simply proposing theories, focusing instead on a rigorous, automated pipeline for generating and quantitatively testing hypotheses against empirical data. Could this framework unlock new insights into the fundamental constituents of the universe and redefine the role of artificial intelligence in scientific discovery?

Unveiling Hidden Patterns: The Challenge of Discovery in High-Energy Physics

Historically, the search for new phenomena in high energy physics has been deeply rooted in the ingenuity of physicists formulating hypotheses and then meticulously testing them against experimental data. This process, while yielding significant discoveries, presents inherent limitations; the formulation of hypotheses is often guided by pre-existing theoretical frameworks and, inevitably, reflects the biases and expectations of the researchers involved. Furthermore, the sheer volume of data produced by modern experiments, such as those at the Large Hadron Collider, far exceeds the capacity for exhaustive manual analysis, creating a bottleneck in the pursuit of novel physics. The traditional reliance on human-driven hypothesis testing, therefore, struggles to efficiently navigate the vast landscape of possibilities, potentially overlooking subtle signals indicative of previously unknown physics beyond the Standard Model.

The Large Hadron Collider and similar experiments now generate data at an unprecedented rate, creating a landscape of possibilities far too extensive for traditional, manual analysis techniques. This deluge of information isn’t merely a quantitative issue; the complexity of interactions within these experiments expands the parameter space-the range of possible variables and configurations-exponentially. Consequently, researchers are increasingly turning to automated hypothesis generation, employing algorithms to sift through the data and identify potentially significant anomalies or patterns that might indicate new physics beyond the Standard Model. These automated approaches aren’t intended to replace human physicists, but rather to serve as powerful tools for exploration, efficiently narrowing the search and highlighting areas deserving of focused investigation within the immense data set.

The observed diphoton counts from the CMS search align with simulated predictions for a pseudoscalar signal coupled to heavy vector-like quarks in both the EBEB (top) and EBEE (bottom) channels.
The observed diphoton counts from the CMS search align with simulated predictions for a pseudoscalar signal coupled to heavy vector-like quarks in both the EBEB (top) and EBEE (bottom) channels.

FERMIACC: An Automated Reasoning Agent for Scientific Discovery

FERMIACC employs deep learning models to create an automated reasoning agent designed to generate potential explanations for observed phenomena. This is achieved through the training of neural networks on existing scientific data and theoretical frameworks, enabling the system to predict plausible relationships and hypotheses. The deep learning architecture allows FERMIACC to move beyond simple data correlation and instead infer underlying mechanisms, effectively acting as an automated scientist capable of proposing novel explanations for complex observations. These generated hypotheses are not based on pre-programmed rules, but rather emerge from the learned patterns within the training data, facilitating exploration of previously unconsidered possibilities.

Hypothesis Generation within FERMIACC constitutes the system’s primary operational function, initiating the proposal of potential explanations for observed phenomena. This process involves defining novel fields of inquiry, suggesting previously unconsidered interactions between known entities, and estimating associated parameters based on both existing empirical data and established theoretical constraints. The system does not simply extrapolate from existing knowledge; it actively constructs new possibilities, allowing for exploration beyond the bounds of current understanding. Parameter estimations are not arbitrary; they are informed by data-driven constraints and adherence to established physical laws, ensuring proposed hypotheses remain plausible and testable.

The AnalysisTemplate within FERMIACC functions as a codified framework for data processing and interpretation. It mandates a standardized structure for input data, specifying required fields, units of measurement, and acceptable data types. This standardization extends to the analytical workflows themselves, defining a pre-defined sequence of operations – including data cleaning, feature extraction, model application, and result formatting – to be applied consistently across all analyses. By enforcing these constraints, the AnalysisTemplate minimizes ambiguity, reduces the potential for human error, and facilitates both automated execution and independent verification of results, thereby ensuring reproducibility and allowing for comparative analysis across different datasets or experimental conditions.

The FERMIAC analysis generator module utilizes a specific schematic architecture to facilitate detailed analysis.
The FERMIAC analysis generator module utilizes a specific schematic architecture to facilitate detailed analysis.

From Hypothesis to Simulation: Modeling the Universe with Computational Precision

UFO Model Construction is the initial step in translating a theoretically proposed hypothesis into a computationally usable format for particle physics simulations. This process involves defining the fundamental particles and their interactions according to the hypothesis, and then expressing these interactions in the form of Feynman rules. These rules are then automatically converted into a Universal FeynRules Output (UFO) model, a standardized file format that encapsulates the Lagrangian and associated amplitudes. The UFO model serves as an interface for numerous event generators, allowing researchers to simulate particle collisions and compare the results with experimental data, thereby testing the validity of the original hypothesis. The standardization provided by the UFO format ensures interoperability between different simulation tools and facilitates collaboration within the high-energy physics community.

Event generation employs the UFO model as input to Monte Carlo simulations, creating a statistically significant number of particle interaction events. These simulations calculate the probabilities and characteristics of various collision outcomes based on the physics encoded within the UFO model, including cross-sections, decay modes, and particle properties. The resulting event samples consist of generated particle trajectories and energies, designed to replicate the conditions and observable signatures expected in high-energy physics experiments like those at the Large Hadron Collider. These simulated events are then used to assess detector performance, optimize analysis strategies, and ultimately, compare theoretical predictions with experimental data.

FERMIACC’s compatibility with existing Effective Field Theory (EFT) frameworks is achieved through a modular design that allows for seamless integration with established theoretical tools and conventions. This linkage enables FERMIACC to leverage pre-existing EFT calculations, parameterizations, and renormalization group equations, avoiding redundant development and ensuring consistency with established results. Specifically, FERMIACC can import EFT models defined in formats compatible with popular EFT libraries, facilitating the generation of UFO models incorporating higher-dimensional operators and their corresponding Wilson coefficients. This capability allows researchers to build upon and extend existing EFT analyses, exploring beyond-the-Standard-Model physics within a well-defined theoretical context and facilitating comparisons with experimental data.

Observed data from the ATLAS dijet resonance plus lepton search aligns with simulated predictions for an octet scalar model proposed by FERMIACC [latex]f597e791[/latex], validating the model's EFT couplings.
Observed data from the ATLAS dijet resonance plus lepton search aligns with simulated predictions for an octet scalar model proposed by FERMIACC [latex]f597e791[/latex], validating the model’s EFT couplings.

Validating New Physics: Statistical Rigor and the Pursuit of Discovery

The interpretation of new physics signals hinges critically on robust statistical analysis, particularly when examining data for anomalies like an excess of high-mass diphoton events – instances where two photons are detected with an unexpectedly high combined energy. Determining whether such an excess is merely a statistical fluctuation or a genuine indication of new particles or interactions requires comparing the characteristics of generated event samples – simulations based on theoretical models – with the observed data. This comparison isn’t simply a matter of matching numbers; it involves assessing the probability of observing the data if the theoretical model were true, using techniques that account for the inherent uncertainties in both the simulations and the experimental measurements. A statistically significant deviation between the generated and observed distributions provides compelling, though not definitive, evidence for physics beyond the Standard Model, prompting further investigation and refinement of the theoretical framework.

FERMIACC distinguishes itself through its capacity to produce a diverse range of hypotheses capable of explaining observed anomalies, moving beyond simply identifying a single explanation. This capability stems from the system’s inherent exploration of the parameter space, allowing it to construct multiple theoretical frameworks that align with experimental data – even if those frameworks differ significantly in their underlying assumptions. Rather than converging on a single ‘best fit’, FERMIACC generates a spectrum of plausible explanations, each potentially representing a valid, though currently indistinguishable, contribution to the observed phenomenon. This approach acknowledges the inherent ambiguity often present in high-energy physics and facilitates a more comprehensive investigation of potential new physics beyond the Standard Model, fostering a richer understanding of the underlying mechanisms at play.

FERMIACC exhibits a remarkable capacity for independent scientific inquiry, autonomously formulating and scrutinizing potential explanations for observed phenomena with a proposal acceptance rate of PASS. This system doesn’t simply generate random ideas; it strategically balances exploration – venturing into novel theoretical spaces – with exploitation, refining promising hypotheses based on existing data. Crucially, this balance is achieved through a ‘model temperature’ of 0.7, a parameter that governs the randomness of the hypothesis generation process; lower temperatures favor refinement of existing ideas, while higher temperatures encourage more radical exploration. The PASS acceptance rate indicates that FERMIACC is effectively navigating this trade-off, consistently producing plausible hypotheses worthy of further investigation and demonstrating a powerful new approach to automated scientific discovery.

The ATLAS diphoton search background and observed counts are overlaid with simulated signal predictions for a hypercharge axion from the FERMIACC proposal [latex]5e41fd9e[/latex].
The ATLAS diphoton search background and observed counts are overlaid with simulated signal predictions for a hypercharge axion from the FERMIACC proposal [latex]5e41fd9e[/latex].

The development of FERMIACC exemplifies a crucial step toward automated scientific discovery, mirroring the ancient pursuit of understanding fundamental principles. As Aristotle observed, “The ultimate value of life depends upon awareness and the power of contemplation rather than mere survival.” This resonates deeply with the project’s core idea of moving beyond mere data processing – the ‘AI Einstein’ approach – toward genuine hypothesis generation and testing. FERMIACC doesn’t simply observe patterns in high energy physics; it actively contemplates possibilities through scaffolded reasoning and deterministic simulations, striving to uncover the underlying logic governing the universe. This echoes the Aristotelian emphasis on active intellectual engagement as the highest form of existence.

Beyond the Search for Elegance

The pursuit of automated hypothesis generation in high-energy physics, as exemplified by FERMIACC, reveals a fundamental tension. Each image, each data point, hides structural dependencies that must be uncovered, but the elegance of a model is a poor proxy for its predictive power. The focus now shifts from replicating existing theoretical frameworks-the ‘AI Einstein’ paradigm-to systematically exploring the vast, largely unmapped space of possible physics. This necessitates a rigorous understanding of the limitations inherent in both the LLM scaffolding and the deterministic simulations employed; biases are not bugs, but reflections of the inductive assumptions baked into the system.

A crucial unresolved problem lies in validation. While FERMIACC demonstrates a capacity for hypothesis generation and testing within the confines of simulation, bridging the gap to real-world experimental data remains a significant hurdle. The next iteration of such models must prioritize not simply finding patterns, but quantifying the confidence in those patterns, and crucially, identifying the conditions under which they are likely to fail. Interpreting models is more important than producing pretty results.

Ultimately, the value of this approach will not be measured by its ability to confirm existing theories, but by its capacity to point toward genuinely novel phenomena. The true test will lie in formulating predictions that, when subjected to experimental scrutiny, reveal something unexpected – a departure from the comfortable symmetries and established narratives that currently dominate the field.

Original article: https://arxiv.org/pdf/2603.22538.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-25 10:12