Current crises challenge both us personally and the economy. While we could laugh at supply chain issues with products like toilet paper and pasta during the COVID crisis, this humor fades in light of the effects of climate change and the Ukraine war. Even if we’re not personally involved in refugee care, the war affects us at the latest when we check our wallets—electricity, heating, and even food are becoming increasingly expensive.
Efficiency – a contribution to sustainability
One consequence of rising costs is the goal of conserving resources: both to cope with sudden shortages in the short term and to sustainably reduce emissions in the long term. This helps us achieve climate goals and ensures a livable world for our children and grandchildren.
Artificial Intelligence (AI) also supports efforts to become more sustainable: at the Excellence Cluster Phenorob, drones and robots are being developed to make agriculture future-proof. But AI is also being used to make processes in the energy industry and manufacturing, the two main sectors of German CO₂ emissions, more efficient.
AI at the beginning of the value chain – opportunities in research and development
A particular opportunity lies in using AI at the beginning of the value chain, that is, in research and development. This area entails high costs and risks for a company: deploying valuable resources is not automatically rewarded with successful development. At the same time, experiments offer the opportunity to try new things on a small scale before mistakes become very costly on a large scale (or with the customer).
In the blog post “How does Design of Experiments work with AI?” as well as in the recently published white paper, it’s discussed how Design of Experiments works technically with AI. But what is Design of Experiments? It’s the art of selecting and conducting only those experiments necessary to answer the current research question from all possible tests. For example, it’s about evaluating with as few resources as possible whether material A is better than material B or assessing with as few (but all necessary) test subjects as possible whether a new drug meets its requirements.
This process typically proceeds in three phases.
The three-phase Design of Experiment process at a glance.
Fundamental to this is defining the problem, i.e., all influencing factors and target variables (Phase 1). This is followed by a preliminary assessment of influencing factors in Phase 2 and optimization in Phase 3. AI can assist in Phases 2 and 3 through informed, probabilistic Machine Learning models: it aids in formulating hypotheses, planning experiments, and analyzing them.
But what is the impact? Application studies show that using AI for Design of Experiments can save up to 50% of experiments. Depending on the setup, this means an equivalent saving of resources in the form of energy, raw materials, and labor time, leading to significantly more efficient and sustainable research and development.
Discussion: What is AI’s resource consumption?
The question is what AI costs in terms of resources when applied in research and development. This is a hotly debated topic, especially for large models that handle large amounts of data. The “bad boy” among Machine Learning methods are language models, for which, back in 2019, the training of a single model in the USA emitted 284 tons of CO₂ equivalents. Even higher numbers are expected for today’s models. By comparison, the average German citizen emits 11.2 tons of CO₂ equivalents per year.
Factors that greatly impact a Machine Learning model’s emissions, and thus its resource and energy consumption, include model type and complexity, the efficiency of the computing cluster, its location (how is the electricity used for operation generated?), and the amount of data used. Sustainable AI research, therefore, requires the targeted use of necessary resources and aims to develop more efficient models.
Regarding AI for Design of Experiments, however, there is reassurance here: typically, small, low-complexity models are used, which aim to extract as much as possible from the limited available data, which is costly. In most cases, all steps can be performed quickly on a laptop—thus, it’s a resource-efficient use of AI and an excellent way to make research and development more sustainable.
Conclusion
In this article, we explored the use of AI to save resources in research and development. AI offers the potential to reduce the number of experiments through smart Design of Experiments while consuming only minimal resources itself. Thus, it can make a relevant contribution to addressing current crises and help us make optimal use of what we have at our disposal.
Further information can be found in the associated white paper:
KI-gestütztes Design of Experiments in Forschung und Entwicklung: Mit Künstlicher Intelligenz und Expertenwissen zu effizienteren Versuchsplänen
Ernis, Gunar und Dorina Weichert, Fraunhofer-Institut für Intelligente Analyse- und Informationssysteme IAIS, 2022, Link
