Simulation and AI/ML riding the digital twin wave
Most organizations across industries have been on overdrive to push digital transformation targeting savings, increasing revenue, or improving quality and reliability. A few common industry challenges such as:
• “Help me predict failure of my plant asset. How do we plan scheduled maintenance optimizing operation, logistics and inventory cost? (OPEX)”
• “How can we maximize performance/throughput from our assets?”
• “How can I send personalize insight and proactive offers to my customers?”
Similar business challenges lie across the industry targeting KPIs along the product lifecycle chain involving design, build and service processes. They can be best approached by scoping and breaking the challenges down to a solution-level design (at Atos, we do this via the 4M-6C methodology). As a matter of principle, without simulating customer challenges, providing any solution would be a futile effort.
Without simulating customer challenges, providing any solution would be a futile effort.
Product designs are systematic, creative and iterative processes requiring physics or mathematical models and manipulated representations. Traditionally, mathematical models have been used to ensure that design fits real-world product usage. During such simulations, engineering models (CAD 3D/2D representations) are engaged, characterizing product and system behavior.
Today’s products are smart, equipped with sensors that continuously communicate about their health. Based on historical data trends and current feeds, modern simulation techniques (artificial intelligence methods) are also able to predict product/system behaviors. They use artificial neural networks (ANNs) or statistical models.
Need for a holistic simulation
AI simulation methods are applied using sensor data with little understanding of products. Constraints such as materials, cost, reliability, performance goals, maintenance and aesthetics are not taken into consideration. Such simulation models become increasingly complex as the number of criteria increases. Moreover, a point of contention remains over how much data is enough for generating a realistic AI model.
On the other hand, physics-based simulation uses assumptions (e.g., data sets), hypotheses and limited external constraints. Such models also involve high-fidelity simulations that continue to grow in complexity and expense, in most cases.
Digital twin bridges best of both worlds
AI simulation methods, based on machine learning, discover hidden trends and deeper correlation and provide insights. They are also capable of modelling and processing nonlinear relationships between inputs and outputs in parallel.
Physics-based simulation has significant advantage, especially when the number of constraints reach beyond certain limits. Additionally engineering constraints such as product geometry, material, it’s kinematics and others help achieve higher accuracy.
A digital twin is a digital replica of a physical entity. They look and react exactly the same way when subjected to similar external conditions. Hence, a true digital twin should represent its physical counterpart through physical representation (CAD model), and behavior based on internal constraints (kinematics) and external stimuli (IoT sensor data).
Bridging the gap between AI/ML and physics-based simulation models produces the best outcomes and accurately represents a true digital replica of a physical entity. This is especially true when real-time sensor data feeds into the physics-based simulations — leveling the playing field with respect to external stimuli.
Applying simulation to digital twins
Adhering to a methodology like 4M-6C helps in scoping and identifying data that addresses specific business challenges. Mapping sensor data to its corresponding physical entities is one of many benefits of this approach. Employing semantics additionally helps to feed time-series data to a physics modelling tool running engineering (design) simulations on metamodels. It’s critical to define thresholds and tune ANN models’ data sets. By doing all things right, ideally, behavior represented by physics simulation should match the neural network-based AI/ML model.
To achieve a true closed loop product lifecycle, this iterative process needs to continue. The AI/ML model must challenge the physics-based simulation to get the most from modern product and process design, which constantly improves over its lifespan.