Autonomous Simulations via Artificial Intelligence for Chemical Processing
Project partners: University of Groningen, Software for Chemistry & Materials BV
Call: PPS-i 2026
Grand date: 28 August 2026
The goal of this project is to develop a fully autonomous, agentic artificial intelligence (AI) framework that allows researchers to interact with computational chemistry and scientific computing software through natural language prompts. The system will translate user prompts into simulation instructions, execute them, analyze the corresponding results, perform error checks, and provide iterative, agentic feedback to the user. By automating these steps, the framework will lower the technical barrier to performing complex molecular simulations, increasing their adoption in both academia and the chemical industry.
Currently, running molecular simulations requires a high level of specialist expertise. Researchers must master the technical details of specific software packages, construct valid input files from scratch, manage job execution, interpret complex output files, and diagnose errors. This can take days or weeks even for domain experts. This steep learning curve severely limits the adoption of computational chemistry in industry, particularly among SMEs that lack dedicated modelling teams. At the same time, even expert users spend a disproportionate amount of time on routine, repetitive tasks that do not require scientific judgement. Our proposed framework addresses both problems simultaneously. It reduces the expertise barrier for new users while freeing experienced researchers to focus on higher-level scientific questions.
The framework will be open-source and modular, built as a mixture-of-experts system in which specialized AI nodes handle distinct aspects of the simulation workflow such as input generation, error correction, and result interpretation. It can be extended to new software packages through dedicated skill documents. This architecture ensures that the framework is not tied to a single simulation program but can be readily adapted to diverse computational tasks, including molecular dynamics (MD), solid-state calculations, and materials simulations. Users will interact with the system through an intuitive chatbot interface accessible via both a command-line interface and a graphical user interface (GUI), making complex computational pipelines accessible without sacrificing flexibility for advanced users.
The development of the AI framework will proceed in three stages. First, real-world data sets, including input scripts, user manuals, and technical documentation, will be collected, curated, and organized into a database. This will feed a retrieval-augmented generation (RAG) system, allowing the AI to retrieve relevant domain-specific knowledge and mimic expert decision-making. Second, model performance will be optimized through prompt engineering, system-level tuning, and iterative feedback, leveraging dedicated frameworks to refine agent behavior without requiring full model fine-tuning, thereby maintaining flexibility and generality while keeping the knowledge up-to-date. Finally, the optimized model will be integrated into an agentic framework capable of end-to-end automation. This will encompass interpreting user prompts, retrieving knowledge, generating and validating input files, executing computational tools, parsing outputs, performing error checks, and presenting results.
The open-source core framework will serve as the foundation for integration into the Amsterdam Modelling Suite (AMS), developed by SCM, one of the consortium partners. AMS is one of the leading computational chemistry platforms in both academia and industry, and is widely used within the Dutch chemical processing sector. By embedding the agentic framework natively into AMS, the project creates a direct and immediate pathway for industrial adoption, supported by SCM’s established software infrastructure and user base.
To validate the framework and demonstrate its industrial relevance, it will be applied to two concrete use cases. The first is the optimization of process parameters for lithium-ion batteries, a problem of direct relevance to the Dutch energy transition. Here, the agent will autonomously execute workflows using the MD modules of AMS to predict lithium-ion transport. The second use case is the prediction of protonation states of oligopeptide and small protein, a key step in computational drug discovery and an important capability for the life sciences industry. In both cases, performance will be evaluated against high-quality reference data and validated through surveys among domain experts, ensuring that results meet the standards required for real industrial application.
The ultimate goal is to provide both the scientific community and the Dutch chemical industry with a robust, openly available AI framework that simplifies complex computational chemistry workflows, accelerates research and development productivity, and serves as a foundational platform for the next generation of autonomous AI systems in scientific computing.