Artur M. Schweidtmann

Piping and Instrumentation Diagrams (P&IDs) are central to process engineering workflows, yet extracting information from them remains a tedious and time‐consuming task. This work introduces ChatP&ID, a framework enabling natural‐language interaction with smart P&IDs through Graph Retrieval‐Augmented Generation (GraphRAG), to our knowledge, the first application and benchmark of GraphRAG to structured engineering diagrams. DEXPI‐encoded P&IDs are transformed into structured knowledge graphs, enabling reliable, grounded querying by large language model (LLM) agents. Benchmarking across commercial LLM APIs demonstrates that graph‐based representations improve response accuracy by 18% over raw image inputs and reduce token costs by 85% compared to directly ingesting smart P&ID files. Among the retrieval strategies evaluated, ContextRAG achieves the highest accuracy (91%) at only $0.004 per query using GPT‐5‐mini. For smaller open‐source models, vector‐based retrieval improves accuracy by up to 40%. ChatP&ID lays practical groundwork for AI‐assisted process engineering tasks, including Hazard and Operability Studies (HAZOP).

Piping and instrumentation diagrams (P&IDs) are central engineering documents whose review remains largely manual, time‐consuming and error‐prone. We investigate generative artificial intelligence (GenAI) for P&ID revision. Specifically, we interpret P&ID correction as a machine translation task. Machine‐readable DEXPI P&IDs are converted with pyDEXPI into attributed graphs, and corrected topologies are represented as generalized SFILES sequences (GGILES). Based on this representation, we adapt a transformer‐based Graph‐to‐SFILES model to utility‐system P&IDs and train it on a synthetic dataset of paired erroneous and corrected diagrams. On this benchmark, the model achieves high accuracy and learns attribute‐dependent error patterns. Tests on five industrial DEXPI P&IDs reveal a gap between synthetic and industrial data, highlighting limitations in data availability and coverage.

Large Language Models (LLMs) can aid synthesis planning in chemistry, but standard prompting methods often yield hallucinated or outdated suggestions. We study LLM interactions with a reaction knowledge graph by casting reaction path retrieval as a Text2Cypher (natural language to graph query) generation problem, and define single- and multi-step retrieval tasks. We compare zero-shot prompting to one-shot variants using static, random, and embedding-based exemplar selection, and assess a checklist-driven validator/corrector loop. To evaluate our framework, we consider query validity and retrieval accuracy. We find that one-shot prompting with aligned exemplars consistently performs best. Our checklist-style self-correction loop mainly improves executability in zero-shot settings and offers limited additional retrieval gains once a good exemplar is present. We provide a reproducible Text2Cypher evaluation setup to facilitate further work on KG-grounded LLMs for synthesis planning. Code is available at https://github.com/Intelligent-molecular-systems/KG-LLM-Synthesis-Retrieval.

Adaptive experimentation and optimization in organic chemistry

Large language model (LLM)-based multi-agent systems (MASs) are a recent but rapidly evolving technology with the potential to transform chemical engineering by decomposing complex workflows into teams of collaborative agents with specialized knowledge and tools. This review surveys the state-of-the-art of MAS within chemical engineering. While early studies demonstrate promising results, scientific challenges remain, including the design of tailored architectures, integration of heterogeneous data modalities, development of foundation models with domain-specific modalities, and strategies for ensuring transparency, safety, and environmental impact. As a young but fast-moving field, MASs offer exciting opportunities to rethink chemical engineering workflows.

Surrogate models are widely used to improve the tractability of process optimization. Some commonly used surrogate models are obtained via machine learning such as multi-layer perceptrons (MLPs), Gaussian processes, and decision trees. Recently, a new class of machine learning models named Kolmogorov Arnold Networks (KANs) have been proposed. Broadly, KANs are similar to MLPs, yet they are based on the Kolmogorov representation theorem instead of the universal approximation theorem for the MLPs. Compared to MLPs, it was reported that KANs require significantly fewer parameters to approximate a given input/output relationship. One of the bottlenecks preventing the embedding of MLPs into optimization formulations is that MLPs with a high number of parameters (larger width or depth) are more challenging to globally optimize. We investigate whether the parameter efficiency of KANs relative to MLPs can be translated to computational benefits when embedding them into optimization problems and solving them to global optimality. We propose a Mixed-Integer Nonlinear Programming (MINLP) formulation of a KAN and test its effectiveness using a case study on the optimization of an auto-thermal reforming process. We observe that KANs are a promising alternative as surrogate models particularly for cases with low number of inputs or outputs. For surrogate models with higher inputs or outputs, stronger formulations must be developed to improve global optimization of models with KANs embedded.