Artificial Intelligence Advances Sustainable Materials Discovery, Addressing Post Synthesis Assessment Inefficiency

Researchers are tackling the critical need for sustainable materials discovery by integrating artificial intelligence with lifecycle assessment. Sajid Mannan (Indian Institute of Technology Delhi), Rupert J. Myers (Imperial College London), and Rohit Batra, alongside Rocio Mercado et al., present a novel framework that co-optimises material performance and environmental impact. This work addresses a significant gap in current materials science, where sustainability is often considered only after synthesis, leading to wasted resources and potentially unsustainable outcomes. By unifying machine learning assisted design with comprehensive lifecycle assessment, the team’s approach promises to accelerate the development of materials that are sustainable by design, rather than as an afterthought, and offers a pathway towards a more resource-efficient future?

This work addresses a significant gap in current materials science, where sustainability is often considered only after synthesis, leading to wasted resources and potentially unsustainable outcomes.

AI for sustainable materials design challenges

Scientists have transformed materials discovery through artificial intelligence (AI), enabling rapid exploration of chemical space via Generative models and surrogate screening. Current AI workflows primarily optimise performance, deferring sustainability assessment to post-synthesis evaluation, which creates inefficiency? By the time environmental burdens are quantified, resources may already be invested in unsustainable solutions. This disconnect between atomic-scale design and lifecycle assessment (LCA) stems from data scarcity, scale gaps from atoms to industrial systems, uncertainty in synthesis pathways, and a lack of frameworks that co-optimise performance with environmental impact.

Researchers propose integrating upstream machine learning (ML)-assisted materials discovery with downstream lifecycle assessment into a unified ML-LCA environment. Recent progress in AI and machine learning (ML) is transforming materials design, evaluation, and deployment. AI-enabled workflows integrating high-throughput Density functional theory (DFT), machine-learned interatomic potentials (MLIPs), and surrogate models now probe chemical and structural spaces previously inaccessible. These approaches have demonstrated success in areas such as battery materials, catalysis, and electronic applications, where data-driven screening identifies promising candidates for synthesis.

Despite these advances, the AI-driven materials discovery paradigm remains performance-centric, prioritising metrics like stability, novelty, and material properties, while performance at the product scale or broader societal and environmental implications are rarely incorporated into early-stage decision-making. This contrasts with mature frameworks in drug discovery, where molecular candidates progress through stages, from target identification to clinical trials, culminating in regulatory approval, with safety, toxicity, bioavailability, and efficacy integrated at every stage. Regulatory bodies mandate this integration, and decades of standardisation have created clear success criteria, validation protocols, and risk assessment frameworks. Materials discovery lacks such a coherent framework.

The workflow remains fragmented: computational screening identifies candidates, experimental synthesis validates feasibility, and characterisation confirms performance, but the pathway from laboratory-scale demonstration to industrial deployment remains poorly defined. Most research on AI-enabled materials discovery focuses on generative modeling of crystals, while a few have synthesised those predicted materials; these represent the first two stages of materials discovery. Unlike drugs, which have defined targets (biological pathways) and success metrics (clinical endpoints), materials serve diverse functions across different application contexts. Furthermore, environmental considerations in materials discovery are treated as auxiliary, post hoc analyses rather than integral design components, an issue exacerbated by AI accelerating materials innovation faster than sustainability considerations are incorporated.

This disconnect between computational generative modeling and real-world materials deployment represents a fundamental gap in the current paradigm. Life cycle assessment (LCA) can quantify environmental impacts across a material’s full lifecycle, from raw material extraction through manufacturing, use, and end-of-life management. By tracking resource consumption, greenhouse gas emissions, toxicity, and ecosystem impacts, LCA enables holistic sustainability evaluation. Conventionally, LCA occurs only after synthesis, characterisation, and pilot-scale production, limiting it to retrospective assessment.

This timing creates inefficiency: by the time environmental burdens are quantified, resources have been invested in potentially unsustainable solutions. Integrating LCA into early-stage discovery would transform environmental impact from post-hoc evaluation to an active design constraint, but achieving this integration is non-trivial. LCA operates at macroscopic scales, industrial processes, supply chains, product systems, while materials discovery operates at atomic and molecular scales. The data mismatch is severe: materials databases contain millions of computed structures, while LCA databases document thousands of industrial processes.

For novel materials, LCA must predict not only properties but also synthesis routes, manufacturing pathways, and in-use behaviour, potentially involving socioeconomic factors such as user acceptance, information inherently uncertain for materials that do not yet exist. This Perspective examines the obstacles to integrating sustainability into AI-driven materials discovery and outlines a roadmap toward ML-LCA frameworks that can co-optimise functional performance and sustainability. Researchers analyse the data challenges, scale mismatches, and uncertainty management requirements inherent to this integration. The goal is not to propose a complete solution, but to chart a course toward discovery workflows where sustainability is intrinsic rather than incidental, where the materials designed for tomorrow’s technologies do not create tomorrow’s environmental burdens.

Materials discovery has traditionally been driven by trial-and-error approaches guided by domain expertise and scientific intuition, where researchers propose candidate materials and validate them through laboratory synthesis and characterisation. The advent of computational methods such as DFT significantly accelerated this process by enabling the prediction of material properties, such as formation energy, band gap, and phonon states, prior to synthesis. This computational screening reduces costs by filtering candidates before expensive experimental validation. However, DFT remains computationally demanding and requires fully specified atomic configurations, limiting its application primarily to virtual screening of known structures from databases such as the Materials Project rather than de novo design of entirely new materials.

Moreover, accuracy remains constrained by approximations in exchange-correlation functionals, and crucial information on synthesizability is often absent. Recent advances in AI and generative modeling, including variational autoencoders (VAEs), diffusion, flow-based models, and large language models (LLMs), are transforming this landscape through inverse design. Rather than predicting properties of known structures, these methods learn distributions of valid material structures and sample candidates conditioned on target properties such as band gap, density, or stability. Complementing generative models, MLIPs enable approximate ab initio molecular dynamics (AIMD) simulations at drastically reduced computational cost, while autonomous experimental frameworks validate synthesizability through closed-loop workflows that propose, execute, and evaluate synthesis routes.

This paradigm promises to unlock previously inaccessible regions of chemical space and reduce reliance on human intuition. Nevertheless, most ML-driven design approaches remain limited to crystalline materials, with few examples for glasses or polymers, and gaps persist in handling novelty, synthesizability, and sustainability regardless of material class. Sustainability, as defined in the Brundtland report, encompasses meeting present needs without compromising future generations’ ability to meet theirs. This concept spans three interdependent dimensions: environmental (conservation of ecological systems and biodiversity), social (human health, wellbeing, and equity), and economic (cost, productivity, and long-term viability).

These dimensions are inherently coupled. For instance, CO2 emissions from materials production contribute to climate change, cascading into ecological damage, social impacts, and economic consequences. Evaluating the ‘sustainability’ of a material thus requires accounting for impacts across all three dimensions. LCA provides the systematic framework for such an evaluation. Unlike conventional materials characterisation that focuses on intrinsic properties, strength, conductivity, thermal stability, LCA quantifies the environmental, social, and economic burdens created by a material or product across its life cycle: raw material extraction, manufacturing, use phase, and end-of-life. Systems thinking is fundamental to LCA: that the impact of a product cannot be assessed from its properties alone but requires accounting for resources consumed and emissions generated throughout its life cycle. Environmental LCA has historically focused on ex-post evaluation, that is, assessing impacts after a product has been manufactured and used.

ML-LCA framework for sustainable materials co-optimisation

The research addresses a critical gap in current materials innovation, where sustainability is often considered after synthesis, leading to wasted resources on potentially unsustainable solutions. This study developed a five-component system designed to co-optimise both performance and environmental impact during the initial stages of materials design. Initially, the team engineered information extraction techniques to construct comprehensive materials-environment knowledge bases, linking material properties with relevant sustainability metrics. Multi-scale models bridged atomic properties to lifecycle impacts, allowing researchers to predict environmental consequences from fundamental material characteristics.

This approach employed Uncertainty quantification to assess the reliability of predicted environmental impacts, crucial for decision-making regarding novel materials. Furthermore, the team implemented uncertainty-aware optimisation algorithms, enabling simultaneous navigation of performance and sustainability criteria. This innovative method allows for the identification of materials that excel in both functional properties and environmental responsibility. These investigations also revealed material-specific challenges that require tailored integration strategies. Realising the full potential of ML-LCA necessitates coordinated advancements in data infrastructure, ex-ante assessment methodologies, multi-objective optimisation, and regulatory alignment.

The work underscores the importance of designing materials for sustainability from the outset, rather than attempting to retrofit environmental considerations after development, ultimately paving the way for a more responsible and efficient materials innovation pipeline., The research addresses a critical disconnect between atomic-scale design and comprehensive environmental impact evaluation, traditionally assessed only post-synthesis. Experiments revealed that integrating these processes upstream, during materials discovery, significantly improves efficiency and reduces the risk of investing in unsustainable solutions. The team measured the propagation of uncertainty from AI models to LCA results, demonstrating a discovery-driven, iterative process where refinement occurs with the availability of new data from experimental validation or improved computational methods. For a candidate material, multiple synthesis routes are predicted, each with probability distributions for key parameters like temperature, pressure, and reaction time.

ML-LCA integrates design and environmental impact assessment

This unified approach addresses the current disconnect between designing materials at the atomic scale and evaluating their environmental impact, which typically occurs after resource investment. The authors acknowledge limitations including data access constraints, the need for interdisciplinary workforce training, and validation strategies for predictive modelling. This work signifies a substantial advancement towards proactive sustainability in materials science. By co-optimising performance with environmental impact from the initial design stages, ML-LCA promises to reduce inefficiency and resource waste. The framework’s ability to quantify uncertainty and navigate complex trade-offs is particularly valuable, enabling informed decision-making throughout the materials development process. Realising the full potential of this approach requires coordinated efforts in data infrastructure, assessment methodologies, and regulatory policy, ultimately contributing to a socio-technical transformation in materials discovery.