In the following, a concept paper is shown for LabIntel, an AI tool to optimize experiments and chemical reactions. The main parts from the concept paper are shown in this article, however the full paper can be downloaded via the following link:

LabIntel is an open source software suite developed under the lead of the Austrian non-profit research center SWISDATA gGmbH. It uses AI to speed up chemical experiments, like finding optimal synthesis parameters for catalysts and for live control of reactions and the application of chemicals. For easier understanding the software’s current and planned features are explained in the following with concrete mostly chemical use cases. Each method can however be adapted and used in different domains, as the implemented algorithms are topic agnostic.

Heterogeneous catalysts are used in nearly all industries to speed up chemical reactions, saving time, resources or even enabling viable application. Synthesizing new catalysts often needs up to thousands of experiments, varying several parameters (mixture ratios, pH value, temperature, timing, oxygen content, …) to find optimal settings. Parameters are often picked at random and experiments are conducted in time consuming brute-force fashion. LabIntel’s collection of custom black box optimization (e.g. bayesian optimization) algorithms can reduce the number of experiments by more than 70%. The researcher first feeds the outline of the experiment into LabIntel, setting up the available parameters, with resolution, precision and range information. The software then recommends a set of experiments that the researcher can try or discard on the spot. LabIntel uses this feedback together with the reported results to recommend additional sets of parameters, guiding the user to find the optimal parameters the quickest.

In this scenario the underlying processes are too complex to understand, e.g., in heterogeneous catalysts or adsorbents, the surface quality can be the crucial factor, which often is affected by effects on molecular or even quantum scale. Thus a black box approach is used, which only analyzes the inputs and outputs but does not fully understand how they are connected.

While other uses cases still are very complex, they might allow more reasoning about the ongoing chemical processes. For example in chemical plastic recycling, plastic feedstock is thrown into a big soup in a reactor. Within hundreds of reactions happen at the same time—many of which are interacting or affecting each other, e.g., byproducts of one reaction might slow down, start or inhibit other reactions. The operator wants to speed up the main reaction, while slowing down possible harmful side reactions. For this more insights of what is actually happening is needed.

LabIntel provides a collection of white box approaches, most notably the reaction finder. It starts from the materials known to be in the reactor and uses a database of known reactions combined with previously learned knowledge to simulate all possible reactions that can occur. This creates a huge tree of possible paths that are consecutively pruned using live sensor data and learns new knowledge about the reaction behavior utilizing reinforcement learning. In the end it delivers possible reactions systems with all reaction constants modeled within confidence intervals. These insights can then be used by operators to understand and optimize their process.

Other white box approaches are regression models, shallow neural nets/perceptrons or found correlation with deeper models using methods from explainable AI like feature attribution.

Both, the white and black box approaches, can be combined into a network of machine learning models. They can be trained online using live sensor data, which allows to find correlations, reactions and to rise the precision and confidence of all trained models. The resulting system of machine learning models forms a digital twin, that can be used to for:

• evaluating what-if scenarios and digitally test changes to the setup
• optimize the runtime for different possible changing aspects, e.g., yield, efficiency, energy use, resource use, and time budget
• find anomalies and alert on possible problems before they happen

The runtime optimization can further be used to optimize a dynamic system. For example in PFAS removal current static systems fail either at short or long chained polymers. With LabIntel a dynamic system can be built, that monitors inflow and automatically tunes all parameters in real-time to adapt to the specific application parameters needed.

LabIntel can practically be used in any process that needs to vary inputs to optimize an output (black box optimization) and for gaining insights and control in many chemical, biological or similar behaving systems (white box approach). Here is a list of use cases currently being tested, or part of proposed projects:

• Plastic waste cleaning: Find optimal synthesis parameters for catalysts using black box optimization. Optimize use of the found catalyst using digital twinning and runtime optimization. Tune the catalyst’s regeneration using black box and runtime optimization.
• PFAS adsorbent synthesis: Optimize pyrolysis of plastic and rubber waste to synthesize adsorbents for most efficient PFAS removal using the black box approach. Tune the application parameters using the white box suite.
• Dynamic system for PFAS removal: Use digital twinning and runtime optimization to tune a dynamic waste water treatment system in real-time for most efficient and resource-saving PFAS removal.
• Chemical plastic recycling: Optimize catalyst design using the black box approach for chemical plastic recycling (similar to depolymerization, but for mixed plastic feedstock). Use the reaction finder to understand and tune the reactions within the reactors.
• Redox flow batteries: Use the reaction finder in reverse, to find synthesis paths for redox active molecules, synthesize and test them in parallel applying high throughput experimentation techniques with LabIntel’s support.
• Bio gas: Use the reaction finder and digital twinning to understand the bio-chemical processes happening within the reactors. Today most operators just dump whatever they have into the tank and hope for the best. LabIntel shines light into the processes, allows to speed up wanted bacterial cycles while slowing down unwanted side cycles/reactions. It further alerts when mass bacteria starvation is evident to happen and with recommendations for possible mitigation.

LabIntel does not try to replace human researchers or operators. Instead it follows the approach of using all available resources (historic data, literature data, AI, statistics, human knowledge) to achieve the best possible result. It works in tandem with humans using their strengths and knowledge. The user interface (UI) is unobtrusive and the user can always intervene—leaving the power and final say to them. For example researchers can already mark large portions of possible input parameters and combinations as likely irrelevant, based on their intuition. LabIntel not only makes this easily possible, but also tries to visualize all data in several ways to provide the user with input they can easily understand and decide on. This makes LabIntel a research and operation assistant that actually helps instead of getting in the way.

A large language model (LLM) based chatbot is directly integrated into the UI. It features SWISDATA’s SmartUI which allows the chatbot to create forms with user inputs, graphs and charts on the fly. This massively reduces training needed to use the software. Users can just ask the chatbot and directly interact with all of LabIntel’s functionality using it. For example the user can start inputting the outline of their experiment by hand and then ask the chatbot when they are unsure how to proceed. The chatbot guides them through the process, asks for all information needed and then inputs it in the application the right way, without further user action required.

LabIntel is open source and is fully built on other open source software. Even for the LLM, only self hosted open source solutions are used. This means that no data is transferred to foreign cloud services and everything can be hosted 100% offline and on premises. This assures stability for the years to come, creates independence and security especially in the face of emerging global crisis and political problems.

Due the open source license, users can make their own additions and adaptions to the software and use it for free. However if needed, professional service contracts and feature development can be provided by SWISDATA.

LabIntel’s development is co-funded by the Austrian government via the Light-AIClean project (see text and image reference below for the project funded by the FFG) . Two more proposals are submitted waiting for a decision, another proposal is currently in the works. As non-profit research center, one of SWISDATA’s core business are public funded projects. SWISDATA will continue to seek public funding and will additionally provide commercial support and services to secure LabIntel’s future.

If you are interested in using LabIntel or in further development, please don't hesitate and contact us. We will answer your inquiry as soon as possible.