This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Obtaining high-quality material data for short-fiber reinforced plastics (SFRP) is a critical yet challenging requirement in material engineering. The mechanical performance of SFRP materials'”commonly used in industries for their lightweight and strength properties'”is highly influenced by factors such as fiber orientation, temperature, and loading conditions. Traditional experimental methods for capturing complex properties like static stress-strain behavior and creep response require significant time, budget, and specialized resources. The delays and high costs associated with gathering this data can be prohibitive, often exceeding the timelines of design projects and limiting the ability to conduct rapid virtual tests, essential for optimizing product performance while reducing environmental impact. The presented work demonstrates a new approach based on pretrained AI models, combined with transfer learning technique, that offers an innovative solution to these challenges. This approach leverages the ability of neural networks that are pretrained on a larger dataset to readapt to a related smaller dataset. The pretrained AI models are created using large experimental material database combined with advanced material modelling simulation data. This allows them to predict complex behaviors such as stress-strain and creep responses, across different dimensions such as temperatures, fiber orientations, and loading conditions. The pretrained AI models are then customized to a new material system with minimal input data using transfer learning. Such approach is particularly beneficial for generating accurate material data where very limited input data is available. The solution demonstrates significant advancements in material data enrichment by simulating the static stress-strain curves and creep performance of SFRP materials under varied conditions. The proposed approach ensures high data fidelity by focusing on limited but crucial input data, which allows for fast, efficient generation of complete datasets without sacrificing accuracy. Experimental validation has been conducted, with results showing strong alignment between AI-generated predictions and actual measured data for both static and creep behaviors. This proves the reliability of the workflow and highlights its value in reducing dependency on extensive physical testing. Moreover, the method provides practical guidance on selecting the minimal set of input data necessary to enrich datasets accurately. For example, stress-strain curves and creep data are enriched across multiple factors'”such as temperature, loading magnitude, and fiber orientation'”allowing engineers to comprehensively model material behavior under real-world conditions with just a few core data points. This work offers a promising alternative to traditional data acquisition for SFRP materials, presenting an efficient pathway to accurate material data generation.