Model-centric data-enabled digital transformation of industrial processes

The Hive
The Hive
Jun 27 · 5 min read

by , Co-Founder and Chief Scientist,

Digital transformation of businesses has received a great degree of attention over the past few years. In a broad sense, digital transformation refers to the pervasive integration of digital technology to enhance organizational performance, leading to new business models and markets. In the services industry, companies such as Uber and Airbnb are noted as harbingers of this transformation, centered around software delivery.

In engineering industries, digital transformations have recently revolved around the concept of a digital twin, which is typically defined as an accurate computational representation of a physical system that uses available sensor information interactively to emulate the operation of the system with the goal of enabling performance forecasts and decisions. Based on the intended application, however, the definition of a digital twin, and its prevalence or lack thereof is widely variable. While it is well-argued that digital twins are being deployed in manufacturing and process control applications, prevailing conceptualizations and implementations remain narrow in scope. To highlight increased opportunities for digital twins, it is worth contemplating two examples:

1. Consider the design of a complex system such as the engine of a commercial aircraft. Detailed simulations of such systems have been employed for over two decades, under the expectation that efficient designs can be realized virtually on a computer and replace physical testing. However, simulations are still not effective enough to drive the design in their own right, as they are too time-consuming and/or inaccurately represent the real-world system. Simulations are instead used to inform the design (in most cases, just to offer trends and insight), along with legacy ideas, empiricism, human intuition, and iterative building and testing of the physical system itself (or sub-systems therein). Further, the design process fails to take advantage of insights from the operational data in a quantitative manner.

2. Consider the operation of complex equipment in an offshore oil rig. Current developments seek to process sensor data using machine learning to optimize maintenance schedules and predict potential failures. The use of physics-based models along with sensor data can improve precision in the decision making and prediction processes. Further, the acquired data is not typically used to directly inform and improve the analysis, design, and manufacture of components of the rig.

The above examples are illustrative of just a few of the existing gaps in virtualization technology, and in the conception of a digital twin. A comprehensive digital transformation can only occur when organizations realize end-to-end twinning, with a focus on the complete product life cycle. This requires a departure from an asset-centric view to one that tightly couples analysis, design, manufacturing, supply chain and operations, along with a broader link to the Industrial Internet of Things. While the notion of tightly-integrated Product Lifecycle Management (PLM) is stressed by most organizations, actual processes tend to be segmented. This level of segmentation exists, for instance, not just between product design and deployment teams, but even within an engineering design team, where disciplinary expertise is embedded in silos and not easily shared in a quantitative manner.

To address these needs, it is trendy to advocate that the upcoming digital transformation should be data-centric. Such an emphasis holds weight from a standpoint that data from every stage of the product life cycle must be managed, curated and transformed into actionable information. In this context, machine learning is proposed as the tool that can deliver intelligence by learning unknown relationships between data, inputs, and decisions. However, this is a limited viewpoint, often promoted from (credible) successes in specific situations. Reliable industrial outcomes require extrapolation to unseen situations and robustness to rare events with the constraint that data may be of variable quality and quantity. Under such circumstances, machine learning-based models are often incomplete, because information content in the data will not be sufficient to accurately capture all relevant details of industrial processes.

A pragmatic approach to digital transformation should be model-centric and data-enabled. Models form the bedrock of the scientific method, and the scientific method provides the foundation for technology that products are built on. In simple terms, models are mathematical descriptions of physical systems, and are typically formulated based on theories which provide a set of rules that govern the behavior of a class of problems. All models are approximations, hence model outputs should be validated using observational data. Going beyond the use of data for validation of models, one can use data to improve the predictive accuracy of models.

In the context of engineering systems, models come in different forms (and nomenclatures): Physics-based models are desirable in analysis and prediction as the underlying hypotheses are typically verifiable. An example is the use of partial differential equations to represent fundamental flow and combustion processes and solve them to compute the detailed airflow inside an internal combustion engine. Phenomenological or Physics-inspired models leverage an understanding of underlying phenomena to directly predict quantities engineers are interested in. An example would be a simple model to provide the output voltage of a lithium-ion battery, without requiring detailed equations as in physics-based models. While the latter may involve more assumptions than the former, they continue to play an important role in design and control because of their lower cost and effectiveness. Data-driven models are primarily driven by machine learning relationships between inputs and outputs and are increasingly in demand with the growth in quality and quantity of data. In addition, domain expertise, intuition, and heuristics are an integral part of product design and operation of complex systems.

Realizing the stated goal of comprehensive digital transformation requires an ecosystem in which all of the above sources of modeling knowledge, data, information and expertise can be seamlessly and interactively merged.

This merger is not simply a matter of orchestrating data flow at a high level using artificial intelligence, but should rather happen at multiple levels in a distributed fashion.

Geminus is developing algorithms and software to provide a platform that enables the creation of these links between models, data and domain expertise. The benefit to an organization is a model-centric, data-enabled digital transformation of the product life cycle by enabling the unification of heavily segmented industrial processes, teams and expertise.

As an example, at a micro-level, the Geminus platform will enable the augmentation of physics-based models for design with machine learning methods. Such a hybrid model offers improved predictive capabilities by leveraging information encapsulated in the data to overcome the impact of physical assumptions while retaining the ability to satisfy known physics. Another realization of the Geminus platform is to use operational data for product design. As another example, field data from a drilling device can be input to a digital twin to recreate its real-world performance and quantify causal relationships that lead to inefficiency or failure. This information can then be directly used to redesign components or to penalize future designs. At a higher level, ensembles of models of variable fidelity and operational data of varying quality can be merged interactively with domain expertise to drive complex decisions.

The key value proposition is precision in designs and decision-making with greatly reduced turnaround times. The model-centric philosophy will also allow for the expression of domain expertise in an interactive and quantitative manner and promote effective collaboration across the organization.

The Hive

The Hive (www.hivedata.com) is a venture fund & co-creation studio based in Palo Alto, CA to co-create startups focused on AI powered applications in the enterprise. The Hive's view is that AI will transform the enterprise and in the long term enable the Autonomous Enterprise.

The Hive

Written by

The Hive

The Hive is a venture fund & co-creation studio based in Palo Alto, CA to co-create startups focused on AI powered applications in the enterprise.

The Hive

The Hive

The Hive (www.hivedata.com) is a venture fund & co-creation studio based in Palo Alto, CA to co-create startups focused on AI powered applications in the enterprise. The Hive's view is that AI will transform the enterprise and in the long term enable the Autonomous Enterprise.