Forecasting and planning optimization tools have long been part of the manufacturing industry’s toolkit. Such tools enable companies to generate action plans based on their knowledge of demand at a certain point in time. Often, however, these plans represent little more than ideal cases relying on events aligning closely with the forecast model. Manufacturers are left unprepared for these unexpected futures with plans that lack robustness against unforeseen shifts in demand, supply, and production.
The challenge is to limit the reduction in performance generated when something doesn’t go as planned or occurs with a variation that exceeds the usual limits. A plan is considered robust if it will lead to good performance in many different conditions.
Digital twins can complete these standard forecasting and prescriptive tools via real-time monitoring of systems. Also, by relying on algorithms, Simulation Digital Twins can reproduce all of these systems' dynamic behavior and have the capacity to predict their state in the future. Simulation Digital Twins make it possible to run virtual experiments to build more robust plans and resilient systems.
Standard Forecasting and Planning Tools Can’t Guarantee Robust Plans.
Uncertainty in today’s manufacturing industry is structural. Companies face uncertainties around future resource availability, lead times, product quality rates, production breakdowns, demand, supply chain, and market volatility which is – for many organizations – approaching unprecedented levels. For example, there is tremendous volatility in forecasts of the future total industry volume in the automotive industry, with some expert scenarios estimating global growth of 10%. In comparison, others forecast a global market reduction of 30%. This uncertainty and volatility require organizations to anticipate disruptions and reinforce the robustness of their action plans. Unfortunately, standard forecasting and planning tools do not empower them to achieve these goals.
Whether a company adopts sophisticated optimization software, embraces a flexible, demand-driven planning methodology, or relies on manual test-and-try explorations based on experience, intuition, and reasoning, the level of robustness achieved remains too limited. When disruptions occur, arbitrary safety margins are insufficient to ensure resilience. These approaches are particularly inadequate when uncertainty about demand, uncertainty about flows (trade barriers), and inter-sectoral effects (scarcity of commodities) combine.
Machine Learning/Deep Learning methods improve forecasting relatively well in stable environments with historical data. However, it quickly reaches its limits in the case of significant or profound structural changes.
However, in digital twins, manufacturers finally have a new tool that can help them escape from production planning perils in an uncertain world.
Four Ways Digital Twins Help Manage Uncertainty
Digital twins, particularly Simulation Digital Twins, empower supply chain planners to manage uncertainty in four key ways.
- Stress testing plans to assess the robustness.
Assessing a plan for its robustness to uncertainty, variability in demand, lead time, and supplier stockouts means gaining visibility on the consequences of both internal decisions and external events on the company. More robust plans help manufacturers avoid overproduction in periods of reduced demand or poor geographic distribution of stock when regional demand shifts.
Spreadsheets and existing optimization tools have difficulty testing operational plans for robustness in the face of uncertainty. They rely on simplified assumptions about the system and neglect the impact of cascading effects. Instead, what is required is a virtual replica of the entire system to represent and simulate the dynamics and step-by-step impact of any disruption across the whole organization and accurately estimate the consequences on performance.
A Simulation Digital Twin offers supply chain managers the opportunity to assess their plans holistically by testing unlimited demand variation scenarios to impact the system. These scenarios can be drawn either from machine-learning-based probabilistic forecasts or from more qualitative sources, including experts within the company. Experiments with Simulation Digital Twins consist of a series of simulations run with random demand. This allows manufacturers to accurately estimate the likely states of the system at a future time, including the expected profit and service levels, the variability of either, or the probability of cash out. While such randomized experiments with a simple model alone might offer some basic information, it’s only when combined with a digital twin that deep value is exposed. In that case, it provides a more granular view, both in time and in space, of indicator variability such as those characterizing bottlenecks. Essentially, the digital twin enables the decision-maker to stress test their system and identify among different plans, the one that is most stable in the face of disruption or that will lead to maximal expected performance over recurrent deviations from the average forecast.
- Directly generate a “robust-by-construction” plan.
Stress testing competing scenarios enables a company to compare those plans and their robustness to demand volatility. Yet, an even more powerful approach is to generate a ‘robust by construction directly,’ that is, a plan that can remain high performing in a large number of different future demand scenarios.
This approach mirrors a strategy adopted by hedge funds in the finance sphere where strategies commonly demand diverse investments that, despite an unpredictable future, are unlikely to fail simultaneously. Hence, an investment in gold holdings to hedge against the risk of inflation has proven a common and robust investment strategy. It is possible to adopt a similar approach to future risk in the manufacturing realm with a Simulation Digital Twin.
Using a technique known as ‘robust optimization by simulation,’ it is possible to account for uncertainties during the optimization process. This technique allows decision-makers to identify optimal stock levels or optimal part sourcing to maximize the average future profit. Days and weeks that fail to meet maximum theoretical profits are balanced by days and weeks where profit is maintained despite significant disruption.
Companies might choose to minimize cash flow risk, minimize stockout probability, or profit volatility instead of maximizing expected profit. Classical deterministic optimization tools cannot achieve such results as they rely on oversimplified models without complex dynamics and random variability. No matter the KPI or combination of KPIs selected, only a digital twin can simulate the complex supply chain dynamics and account for cascading effects and variability to deliver a robust plan.
- Identify risks, monitor variations, and create alerts.
Even the most robust plan is subject to some fragility. For example, regional demand for a product can suddenly and drastically diminish, or lead times may increase rapidly and significantly. If these variations - in combination or otherwise - are too extreme, even a moderately robust plan may not be able to absorb the immediate or delayed consequences effectively. Contingency planning -identifying risks, their probabilities, and their impacts- is essential, but developing such plans with standard tools and processes is frustrating. Determining the best contingency plan can prove just as problematic as determining the optimal production plan in the first place. Extreme events push organizations with too-rough plans into danger as the simplifying assumptions with arbitrarily defined security margins prove problematic in the real world. Traditional tools cannot consider nonlinear, threshold, and cascading effects that typically reveal themselves in these cases, sometimes even leading to company cashouts.
A large-scale virtual sensitivity analysis, for example, can automatically identify which types of variability are the greatest threat and the point at which they become most disruptive. For example, this sensitivity analysis might identify that a lead time increase of 10% or 20% is manageable, but an increase of 30% will spread major problems elsewhere in the enterprise. Armed with this information, companies can run additional virtual experiments to create adapted contingency plans ready to activate should lead times, or any other variable, prove detrimental to the overall operational plan.
In effect, the digital twin will have allowed the organization to build a robust plan, identify the limits of that plan and the points at which it is most fragile, and prepare contingency plans should any of those points be breached.
- Build autonomous adaptive strategies for real-time decisions
Creating autonomous decision-making systems that automatically adapt in real-time to any situation takes robust planning to the next level. For example, such a system would enable a manufacturer to adjust stock levels in light of current lead times, product quality, and short-term trends in product demand to maximize the final profit. This autonomous decision-making is reactive in real-time to the operational system and can adapt automatically to the organization and its changes.
Manually defining such strategies or decision-making rules is difficult because it demands the capacity to define what should be decided in real real-time and all situations. However, by combining deep neural networks and a Simulation Digital Twin, a technique called ‘deep reinforcement learning’ helps uncover the best strategy to adopt. In much the same way that a champion tennis player has learned how to best return a ball in almost all (including unknown future) situations by training for hours upon hours, the algorithm learns through millions of virtual experiments in the digital twin how to best react in any given situation.
Once trained on the simulation digital twin, such smart decision algorithms can be deployed to control the real system in real-time. Drawing on current data (for example, the location of a part, lead times, or demand shifts), the algorithm can help improve an organization's resilience by making it more agile, flexible, and instantaneously reactive to new never-before-encountered demand curves.
Conclusion
Standard forecasting and planning tools offer only best guesses and best case scenario plans for organizations based on historical data, industry experience, or – at best – simplified models of an operational system. On the other hand, Simulation Digital Twins help supply chain managers gain visibility of their entire system and generate less sensitive or more robust plans. Operators can then better anticipate through monitoring and virtually testing contingency plans by simulation so they can react in real-time to variations with autonomous decision-making strategies that are both flexible and reactive.
About the Author
Pierre-Alexis Gros
Chief Scientist, Cosmo Tech