Future of machine learning in geotechnics

TLDR

Machine learning is widely used across industries, sparking interest in applying it to geotechnical engineering. This paper proposes a data‑centric agenda for geotechnics, advocating that future ML should prioritize data, novel algorithms, and practice relevance. The authors outline a data‑centric geotechnics agenda comprising data centricity, practice fit, and geotechnical context, and highlight data‑driven site characterization as a key research focus. Challenges identified include handling poor data, achieving explainable site recognition, ensuring ML indispensability, meta‑learning, and developing digital twins.

Abstract

Machine learning (ML) is widely used in many industries, resulting in recent interests to explore ML in geotechnical engineering. Past review papers focus mainly on ML algorithms while this paper advocates an agenda to put data at the core, to develop novel algorithms that are effective for geotechnical data (existing and new), to address the needs of current practice, to exploit new opportunities from emerging technologies or to meet new needs from digital transformation, and to take advantage of current knowledge and accumulated experience. This agenda is called data-centric geotechnics and it contains three core elements: data centricity, fit for (and transform) practice, and geotechnical context. The future of machine learning in geotechnics should be envisioned with this “data first practice central” agenda in mind. Data-driven site characterization (DDSC) is an active research topic in this agenda because an understanding of the ground is crucial in all projects. Examples of DDSC challenges are ugly data and explainable site recognition. Additional challenges include making ML indispensable (ML supremacy), learning how to learn (meta-learning), and becoming smart (digital twin).