Athanasios Kontopoulos, Computational & Data Science Scientific Director, R&D, and International Fellow, Air Liquide
IoT is bringing data. AI is there to make a competitive advantage out of it. AI is galloping with innovations, and the R&Ds of the big, medium, or small enterprises can be its spearhead. However, R&D works mainly on technologies. We will explore the conditions needed to succeed, as well as some AI applications and future trends.
What is needed to succeed?
What we first learned is that we need an internal and external ecosystem to make it happen.
To start with, staying close to the business is the cornerstone. The businesses are to be integrated into the whole process, from the very beginning until the final implementation and scale-up.
Then, a “3-D” approach is interesting to apply:
● Skills - people. Technical skills are essential, of course. R&D usually focuses on data science (in Air Liquide, R&D has currently the biggest team of data scientists). However, some other skills are a must, such as IT architecture and programming, data engineering, design, which are skills that mainly lie within the Digital & IT department. Soft skills also are essential: training, communication,...
● Strategy and governance: the evolution of the Organization is key. We defined functions as data owners, data officers, and central data governance bodies. Because if no (good) data, no AI!
● Infrastructure. Data needs to be stored in a structured and harmonized manner, in the cloud, and accessed easily.
"Staying close to the business is the cornerstone. The businesses are to be integrated into the whole process, from the very beginning until the final implementation and scale-up"
For data strategy, governance, and infrastructure, the Digital&IT Department is responsible, within Air Liquide, working hand-in-hand with Operations and R&D. This is the internal ecosystem needed to allow data access, on the one hand, and scale-up on the other hand. If not, R&D could easily get stuck in performing proofs-of-concept that are difficult to put in industrial practice.
We also foster collaborations with many industrial partners and top-tier Universities and Institutions in Europe (CentraleSupélec, Mines ParisTech, Fraunhofer,...) and in the USA (Carnegie Mellon, Wharton,..) as well as with small innovative companies. This external ecosystem is another condition for success.
Some applications, technologies, and future trends
The framework we use in Air Liquide for our digital transformation is: Assets-Customers-Ecosystems (ACE). We will give just some examples.
● Let us take one common application of AI: predictive maintenance. In Air Liquide, we have deployed it on a worldwide basis through our “Smart Innovative Operations” program. Although very efficient, what is usually meant by the term “predictive maintenance” is anomaly detection. The technology behind it is, most often, a clustering algorithm. The future trend is to move towards prescriptive maintenance and maintenance policy optimization. From a tech perspective, we would need a combination of statistics, machine learning, and physical modeling to get there. We are doing R&D on it within a collaborative program with other industrials and academics to increase our chances of success.
● Digital twins are already used in our Operations, mainly to optimize decisions in real-time. The future trend here: combine physical modeling and AI in order to take advantage of the precision of the first and the speed of the second in order to make the design of our units faster. Again, we are leveraging our R&D ecosystem, working in a related program with many industrial companies and academics.
● One of the many applications developed is a bundle and cross-sell for affiliates that are in the packaged gases and hardgoods business. This needed R&D to work on the algorithms (matrix factorization) to adapt them to a BtoB environment. The scale-up of the application requested, of course, close collaboration with the IT and the Operations. One future trend here: test what deep learning can bring, balancing potential benefit in accuracy vs. complexity.
● Here we can mention energy transition/hydrogen economy topics. As renewable energies are expected to increase their part in the mix and hydrogen being part of the game, we would need to understand better operations and economics of greater ecosystems, such as basins. We are already working on the topic, but R&D will be intensified, leveraging academic partnerships, using financial mathematics, complex energy systems modeling, and optimization under uncertainties, together with machine learning.
● We need more and more to perform calculations efficiently on data coming from many different parties while still maintaining confidentiality in these larger ecosystems. Thus testing-related technologies (homomorphic encryption, blockchain,...) is an important R&D topic for the next few years.
Some transverse topics:
● As AI systems become omnipresent and more complex, ensuring their reliability and robustness as well as better interpretability is necessary. To do so, we recently integrated a vast collaborative R&D program with many industrial and academic partners to ensure a trustworthy AI. It is also aligned with Air Liquide’s strategy to have our AI systems complementing and augmenting the human, putting the human at the center.
● To make machine learning more accessible and faster to obtain results with, we will be exploring “autoML” techs (automatic feature engineering, auto-encoder,...)
As a conclusion, R&D can bring exciting new AI applications in manufacturing, but working hand in hand with an internal and external ecosystem is a “must” to ensure value creation.