Artificial intelligence has changed our lives as consumers. Why hasn’t it changed our industries? AI in industry requires more than just big data to work, and the solution lies in the world of physics.
If a predictive algorithm fails in the consumer industry, it’s not the end of the world. Maybe an ad doesn’t get clicked or a TV show doesn’t get watched.
In asset-heavy industries such as oil and gas, power and utilities and manufacturing, failure isn’t an option. A poorly trained algorithm could cause production to grind to a halt, damage equipment, or — in the worst-case scenario — put lives at risk.
Despite the high-value potential of AI in industry, there are few success stories to tell. This is because of the fundamental differences between the classic applications of AI, such as natural language processing, advertising and games — and those used for industrial problems.
Hybrid AI, the combination of physics and data science, is the key to unlocking the potential of artificial intelligence in industrial settings. AI finds patterns from information, and by adding physics to the mix, we provide more information — and more importantly, information that’s accurate enough to power solutions that create value and accelerate the digital transformation of industry.
The ImageNet database contains more than 14 million pictures for visual object recognition. The amount of text available for natural language processing is almost unfathomable. For some applications, it’s even possible to automatically generate training data.
No such resources exist for industry. Although a typical oil platform or factory will be instrumented with thousands of sensors that may have been collecting data for decades, the actual amount of relevant data is often small due to operational changes in industrial equipment over the years. In addition, there may be little to no data for the equipment in its optimal operating zone, and traditional AI models may produce inaccurate predictions when extrapolated outside the existing data range.
Data quality is another challenge that separates industrial AI from classic AI. The physical sensors at industrial sites are often located in harsh environments, which means the data is subject to varying degrees of noise, bias and different raw compression levels. In comparison, a typical data set used by classic applications of AI will have no or negligible noise levels.
However, there’s one area where industry has an advantage. While there are few mathematical models describing consumer behavior, the majority of the problems in industry are governed by the laws of physics and can be described using mathematical and phenomenological models that form the basis of advanced simulators. Industrial subject-matter experts have been using these simulators for decades to support critical decisions. The disadvantage of simulators is that they are often computationally expensive compared to purely data-driven models and may not be feasible for real-time predictions of the variables of interest.
Physics simulators and AI models clearly complement each other. The former can make predictions about future events and outside the range of data used to create and validate models, while the latter can be set up without any knowledge of the underlying physics and work even on a small set of sensors. Combining the two methods keeps the strengths and reduces the weaknesses of both.
How To Get Started With Hybrid AI
It can be challenging to know how best to deploy hybrid AI solutions at scale, given that every business and industry has specific requirements. Hybrid AI is best suited for complex industrial process problems where a mathematical theory framework exists. To make industrial hybrid AI work, I see the following as prerequisites:
• Understand The Problem: The accuracy of hybrid AI depends on the use case or problem it’s solving. Leverage the in-depth knowledge of your subject-matter experts to understand specifics.
• Implement The Tools Necessary For Feature Engineering: These tools allow you to quickly extract value from your data with little effort, time and cost. While some tools are simple and can be implemented by your internal IT department, others require professional intervention.
• Ensure Sufficient Access To Physics Simulators: Choose whether to implement an on-premises solution or invest in a simulator-as-a-service from a reputable vendor. The advantages of service arrangements include reducing initial expenditures and the ability to deploy immediately and at scale.
Industrial companies can’t skip ahead to this part of the digitalization journey, however. The first step must always be data contextualization. Prioritize data organization over centralization. Start driving the connection and mapping of all relevant data sources with a clear list of target use cases in mind. Only then can you scale to include hybrid AI, digital twins and other industry application suites — and later build data products around them.
Hybrid AI Requires An Industrial DataOps Mindset
Hybrid AI starts with Industrial DataOps. One of 2021’s top software trends, industrial DataOps is all about breaking down silos and optimizing the broad availability and usability of industrial data. But beyond just data accessibility and contextualization — which are important — industrial DataOps is a practice. It’s a way of engaging and collaborating across the organization to both share and reap greater value from data.
Industrial DataOps is the starting point to solving one of the key challenges of industrial digitalization: enabling hybrid AI solutions at scale. Only by evolving your organization’s data mindset and creating one central, contextualized source of truth can you ensure that all necessary live data is available for both hybrid AI models and the subject-matter experts supervising the decision-making processes that optimize production and create value.