Optimising Inventory with AI: From Static Models to Adaptive Systems
Getting inventory allocation right across multiple warehouses is a classic problem, and a stubborn one.
Too much stock in one location, not enough in another. Orders placed too early, or too late. Transport costs creeping up while service levels drop.
A new platform developed by MIT’s Centre for Transportation & Logistics and Mecalux, called GENESIS, takes a different approach. Instead of relying on static rules or single-model forecasts, it combines optimisation, simulation, and machine learning to explore a wide range of possible strategies before recommending a course of action.
From Forecasting to Scenario Exploration
At its core, GENESIS ingests inputs such as regional demand forecasts, transport costs, and warehouse capabilities. But rather than producing a single “best” answer, it evaluates thousands of potential distribution and replenishment scenarios.
One of its key features is the ability to recommend internal stock rebalancing. Instead of triggering new orders when inventory runs low, the system can suggest moving stock between sites, improving utilisation and delaying unnecessary procurement.
That shift, rom reactive ordering to system-wide optimisation, is where much of the value lies.
Combining Optimisation, Simulation and ML
What makes this approach interesting from a data science perspective is the combination of techniques.
Genetic algorithms are used to search across large solution spaces. Machine learning models help estimate demand and system behaviour. Simulation allows those decisions to be stress-tested under different conditions.
Together, they form a loop:
- generate potential strategies
- simulate outcomes
- evaluate performance
- refine and repeat
This kind of hybrid modelling is becoming increasingly important in complex, real-world systems.
Faster Decisions, Better Trade-offs
According to the research team, analyses that once took days can now be completed in minutes. That speed enables more frequent scenario testing and faster responses to changing conditions.
The platform also models trade-offs beyond simple inventory levels. It considers decisions such as:
- Whether to consolidate shipments or fulfil from specific nodes
- How to balance lead times against transport costs
- How demand variability affects stockout risk
Results are presented through dashboards that translate model outputs into something usable across teams, not just by analysts, but by operational decision-makers.
What Comes Next: Digital Twins and Warehouse Optimisation
The next phase of development extends beyond inventory allocation.
Plans include:
- Modelling internal replenishment routines
- Building digital twins of automated storage systems
- Exploring slotting optimisation to improve picking efficiency
This points towards a broader shift, from isolated optimisation problems to integrated, system-level modelling.
Decision Support, Not Decision Replacement
Importantly, GENESIS is positioned as a decision-support tool.
Its outputs are recommendations, not instructions. They are designed to be interpreted alongside commercial priorities, operational constraints, and human judgement.
That distinction matters. Especially in environments where trade-offs are complex and context-dependent.
Why This Matters for Data Scientists
This kind of system reflects a broader trend in data science.
It’s not just about building models anymore. It’s about designing systems that:
- Integrate multiple techniques
- Operate under real-world constraints
- Support ongoing, adaptive decision-making
For those working in supply chains, logistics, or operations, this is where a lot of innovation is happening.
And more broadly, t’s a reminder that some of the most interesting data science problems aren’t just about prediction.
They’re about making better decisions in complex systems.