Frugal AI in geophysics

Development of methods of artificial intelligence in geophysics often require heavy computations using high-performance computing (HPC) with graphics processing units (GPUs). Given the ongoing energy crisis and exponentially growing demand for data storage and computational resources, the concept of Frugal AI has emerged recently. While there are attempts to use renewable energy for supplying data centres and HPC clusters, globally the energy supply remains being dominated by fossil fuels [1].

Frugal AI proposes to analyse and optimise the use of energy and water resources for data centres, computational power for data processing, AI model training and deployment for environmental sustainability of AI.

In support of the EU AI Act [2], there are significant efforts to develop standardisation documents for environmentally sustainable AI solutions (see draft technical reports ISO 20226 [3] and CEN/CENELEC 18145 [4]). The purpose of these standards is to offer AI stakeholders a practical approach to quantification of the environmental impact of AI solutions. In particular, carbon emissions of scopes 1,2,3 can be quantified with associated uncertainties. Scope 1 are direct emissions [5] which may occur in data centres that use reserve fossil-fuel generators (this is usually a temporary solution for autonomous power supply). Scope 2 are indirect emissions that are produced due to electricity consumption. It is possible to quantity such carbon emissions using a dynamic carbon factor based on real-time fuel mix in electricity generation, known power of used hardware modules (processors, memory modules, and GPUs), and duration of training and deployment of an AI solution. Scope 3 carbon emissions [6] are based on life-cycle assessment of an AI solution (“from cradle to grave”), which includes hardware manufacturing and software development and deployment.

It is not seldom that AI models are deployed in default non-optimised mode for trial-and-error experiments, which may take a lot of resources and produce a large carbon footprint. The output data of large-scale geophysical models augmented with AI may be stored in an inefficent way, which can be improved using good practices (in appropriate formats, suitable temporal and spatial resolution).

We will discuss various aspects of Frugal AI in geophysics and suggest what optimised techniques can be used at each stage of AI development for geophysical applications. Frugal AI development may include simpler algorithms that can achieve comparable results with significantly smaller energy consumption; refined processes of data gathering for capturing essential information for AI training (for example, sparse datasets); optimised designs of training models; and deployment of AI models for energy-saving technologies, such as Demand-Side Response (DSR). Further intervention may include use of energy-saving hardware units and analysis of AI life cycle, which would identify stages of AI use that require most resources, and how those can be reduced.

References

[1] World energy report 2023, https://www.energyinst.org/__data/assets/pdf_file/0006/1542714/684_EI_Stat_Review_V16_DIGITAL.pdf 

[2] The EU Artificial Intelligence Act, https://artificialintelligenceact.eu/

[3] ISO DTR 20226 Information technology --- Artificial intelligence --- Environmental sustainability aspects of AI systems.

[4] CEN/CENELEC DTR 18145 Environmentally Sustainable AI

[5] ISO 20181 Stationary source emissions. Quality assurance of automated measuring systems.

[6] ISO 5338 Information technology --- Artificial intelligence --- AI system life cycle processes.