Session: 34-01 AI for flow field prediction and post-processing

Submission Number: 176785

Harnessing AI for Scalable Analysis of Simulation Data 

High performance computing allows engineers to generate simulation data at ever-increasing rates. Whether generated by running 100s of steady-state simulations (design optimisation or design space mapping) or a small number of time-resolved computations (full-annulus unsteady, or high fidelity solvers), this data provides a challenge for storage and analysis. The amount of persistent storage required is typically reduced by only retaining subsets, often called “extracts”, of the full domain. These extracts can then either be viewed statically as individual cases or dynamically using interactive filters on high-level meta-data (key input and output metrics of each case) to select, synchronise and display views of chosen cases in real time; the dbslice visualisation framework is an example of the second, dynamic approach.

In this paper, dbslice is extended to harness the capabilities of large language models (LLMs). An important system design decision is that the server provides access not only to the data (meta-data and extracts) but also to common tools used to process the data. In this way, we ensure that the analysis requested by the LLM is traceable, reusable and verifiable. The data and tools are exposed in two ways: a web API and a Model Context Protocol (MCP) server. The web API allows remote access and also provides an endpoint for natural language chat applications. The chat endpoint: receives a question from the user as input; uses an LLM to formulate a multi-step plan to answer the question using dbslice tools; executes the plan; and, returns an answer or document, generated by an LLM, using the results of the foregoing analysis. The MCP server offers the same tools via the emerging MCP standard that allows LLMs to connect with data services, thus allowing broad interoperability.

The approach is demonstrated on a database of 590 axial compressor simulations. By asking a question about the influence of compound lean, a multi-page report is generated that provides correlations between lean and endwall loss (providing statistics and scatter plots) and identifies regions of the extract images that are influenced by the lean parameter (highlighting corner separation control). The report also links the current analysis with the findings of curated set of technical publications, providing guidance and recommendations.

The framework described in the paper, combining access to both data and verified tools, provides a scalable strategy for engineers to harness LLMs in the analysis of their data.

Presenting Author: Graham Pullan University of Cambridge

Presenting Author Biography: Professor Graham Pullan is Co-Director of the Whittle Laboratory at the University of Cambridge. He leads the EPSRC Centre for Doctoral Training in Future Propulsion and Power. He completed his PhD in 2001, at the Whittle Laboratory, under the supervision of John Denton. His research is focussed on the three-dimensional aerodynamic design of turbomachinery, and on the development of the computational methods required to achieve this. He has received: the 2012 ASME Gas Turbine Award; ASME IGTI Best Paper Awards in 2011, 2012, 2014, 2016, 2017 and 2018; the 2013 Rolls-Royce Howes Ruffles prize.

Authors:

Graham Pullan University of Cambridge