The growing scale of artificial intelligence (AI) and machine learning (ML) has sparked public discussion about their energy use and environmental impacts — especially during training and deployment of large, deep-learning models. Existing approaches, such as Bayesian optimisation, evolutionary algorithms, and pruning or quantisation alone, provide minimal energy savings and do not account for multiple optimisation stages. This paper presents a Quantum-Inspired Algorithm (QIA) for sustainable machine learning that overcomes these limitations by jointly optimising hyperparameters, model compression, and scheduling in a single framework. To answer that, QIA uses tunnelling-based search and amplitude reweighting to find energy-efficient configurations faster, and builds a metric comparison to evaluate trade-offs among accuracy, energy consumption, and carbon footprint using a new Quantum-Sustainable AI (Q-SAI) index. QIA reduces training energy consumption by 20–30% compared to several baseline optimisation methods while keeping accuracy losses under 1.5% across experiments on CIFAR-10, ImageNet, IMDB, and GLUE. Optimised inference latency and per-query energy consumption are demonstrated, along with tangible reductions in carbon footprint in deployment regions. Our results show that QIA offers a scalable, domain-agnostic solution for embedding sustainability considerations into contemporary machine learning workflows.
Reddy et al. (Tue,) studied this question.