The pursuit of scientific discovery traditionally fuses observation, analysis, and formulation of new hypotheses, but researchers now explore how machine learning can accelerate this process. Maximilian Negele and Florian Marcardo, together with Max Planck Light of Light and Erlangen Nuremberg's science colleagues at Friedrich Alexander University, present a new approach to fully automating scientific quest. They introduce SciexPlorer, an agent that utilizes the capabilities of large-scale language models to investigate physical systems without relying on pre-programmed instructions or task-specific blueprints. This innovative agent successfully explores a range of models, encompasses mechanical dynamics, wave evolution, quantum physics, demonstrates its impressive ability to recover fundamental equations, discover important properties from observed data, and hosts automated methods of scientific discovery in multiple disciplines.
Agent Framework for Physics Model Exploration
Agent exploration of physics models offers a new approach to scientific discovery, moving beyond automation of specific tasks to authentic scientific institutions. This work introduces an agent-centric framework that repeatedly proposes experiments, interprets results, and improves understanding of a particular model. The agents operate within a closed-loop system, interact with a simulated environment representing the physics model under investigation, and use Gaussian process regression to construct predictive models, allowing efficient investigation and prediction of experimental results. Importantly, agents incorporate inherent motivational mechanisms to reward rewards to reduce prediction uncertainty rather than achieving specific task goals.
This intrinsic motivation causes agents to actively seek beneficial experiments even in the absence of external rewards. The team demonstrates the effectiveness of the framework in a simplified model system, showing that agents can autonomously discover autonomous relationships between model parameters and observable quantities, and successfully identifies hidden parameters that manage system behavior with 7% accuracy. This represents a critical step in creating an artificial scientist capable of independent scientific research and potentially accelerate the pace of discovery in complex systems.
Automating the open-ended iterative loops required to discover unknown systems' laws through experiments and analysis remains a critical issue. Here, the team introduces SciexPlorer. SciexPlorer is an agent that leverages the large-scale language modeling tool usage capabilities to enable free-form exploration of systems without domain-specific blueprints. They applied Sciexplorer to search for physical systems that were initially unknown to agents, tested on a wide range of models spanning mechanical dynamic systems, wave evolution, and quantum many-body physics, showing promising results in automated scientific exploration.
LLMS accelerates discovery in the scientific field
Recent research details the surge in the application of large-scale language models (LLMs) such as GPT-5 and Gemini to accelerate scientific discovery in a variety of fields, including chemistry, materials science, biology, physics and engineering. LLM is used to predict chemical properties and reactions, design new materials with specific properties, and simulate material behavior. In biology and biomedical, LLM is applied to biomedical text mining, protein structure prediction, and drug discovery, but within physics, it is used to govern equations from data, analyze complex physical systems, and automate simulations. A notable trend is that LLM combines with tools such as simulators and databases to provide the autonomy to plan and execute scientific tasks, allowing closed-loop experimentation and discovery.
While generic LLMs are useful, fine-tuning them with domain-specific data can greatly improve performance, as shown in special models such as Biobert and BioGpt. Several frameworks, including Sciexplorer, Clapp, and open source planning and control systems, are being developed to streamline LLM integration into scientific workflows. Jax, a high-performance numerical library, is used to scientific simulations and build models, and many of these projects are open source and can be used on platforms such as GitHub to promote collaboration and reproducibility. However, challenges remain, such as data availability and quality, ensuring reproducibility and verification, and improving interpretability and explanationability. Scaling and generalizing new issues also present ongoing challenges, as well as seamless integration with existing scientific infrastructure. This study highlights a paradigm shift in scientific discovery, which has become an increasingly powerful tool for automating tasks, generating hypotheses and accelerating the pace of innovation.
Autonomous scientific discoveries through agent exploration
Sciexplorer represents an important step in automating scientific discoveries through the development of agents that can independently explore unknown physical systems. The team successfully demonstrated that the agent could leverage the capabilities of large-scale language models and code execution to infer motion and Hamiltonian equations from observed dynamics without prior knowledge or task-specific instructions. SciexPlorer autonomously generates and executes Python code to extract qualitative signatures from data, builds candidate models, fits into observed accelerations, and effectively replicate aspects of the scientific process. Agents achieved strong performance in a variety of mechanical, dynamic, and wave-based systems, as measured by measurement coefficients between predicted and actual dynamics, and often recovered governance models with high accuracy. Although systems are excellent at identifying systems similar to those within existing knowledge bases, performance decreases in the face of completely novel or complex scenarios, highlighting the dependence on existing knowledge. The authors acknowledge that the general mode of failure includes an early commitment to false models and a limited ability to reconsider early assumptions when faced with poor fits, and that future work may focus on improving agents' capabilities for self-correction and expanding their ability to explore truly unexplored science fields.
