Decolonising AI in materials science: Structural patterns and interpretability gaps in atomic discovery models
Original framing: “Interpretable AI in materials discovery: Uncovering how models make predictions” — Phys.org
The history of material science is intertwined with colonial extraction, from the transatlantic slave trade's demand for dyes to the lithium boom in the Global South. Western material ontology emerged from Enlightenment-era reductionism, which severed materials from their cultural and ecological contexts. The current AI-driven approach echoes 19th-century positivism, where data is treated as neutral while power structures dictate what counts as 'discovery.' Parallels can be drawn to the Green Revolution, where technological solutions masked structural inequalities in agriculture.
The AI method for interpreting material discovery is framed as a technical breakthrough, but its roots lie in a colonial, reductionist paradigm that treats materials as inert commodities.
Indigenous epistemologies—such as Māori *whakapapa* or African *Ubuntu*—offer relational frameworks where materials are active participants in ecological and cultural systems, challenging the atomistic assumptions of AI models. Historically, material science has been complicit in extractive industries, from colonial mining to the lithium boom, and the current AI approach risks perpetuating these harms by prioritising 'efficiency' over regeneration. Marginalised voices, including Indigenous scientists and Global South researchers, are systematically excluded, reinforcing epistemic violence. A systemic solution requires decolonising datasets, centring regenerative design, and establishing community-led innovation hubs that redefine 'usefulness' through ecological and cultural integrity. The trickster figures of Hermes and Eshu remind us that the solemnity of 'progress' obscures the absurdity of treating life as data—until we invert the paradigm entirely.