The accurate characterisation of the elasto-plastic response of materials is important for ensuring the reliability, safety, and performance of various engineering designs and structures. Conventional characterisation methods often rely on specialised experimental setups and inverse identification procedures, which can be time-consuming and frequently destroy the material specimen. This work aims to establish a new approach for determining the elasto-plastic response of metallic materials by combining scratch testing with machine learning (ML). Scratch testing is a minimally destructive experimental approach mainly used for evaluating the mechanical and tribological properties of surfaces. In the literature, only limited morphology information from the scratch test is used for analysis, such as groove pile-up height, width, and residual depth [1,2]. In our novel approach, additional morphology features are extracted from the residual scratch morphology to allow for characterisation of additional material parameters. Finite element modelling is the standard approach in the literature for generating data for analysis [1,2]. However, we propose using the Halton Sequence and dynamic sampling to provide better coverage of material parameters than traditional grid sampling and generate a more diverse dataset. The inverse problem is addressed using ML. Some success has been found using multilayer perceptrons (MLPs) [3]. However, building upon this, extreme gradient boosting (XGB) and symbolic regression (SR) are used. A primary focus is placed on SR, as this allows for determining parsimonious semi-analytical expressions that relate the residual scratch morphology to the elasto-plastic material parameters. Additionally, it is expected that SR will generalise better than other ML methods, leading to more versatile models. Overall, the proposed framework offers an alternative approach to elasto-plastic material characterisation by combining scratch testing and data-driven modelling.