Industrial AI optimisation tools depend on large historical datasets and long retraining cycles, but these datasets are often unavailable in the production environments that need optimisation most: recycled and biobased plastics, frequent changeovers, novel formulations and variable material batches. As processes change faster than data can be collected, manufacturers face material waste of up to 30%, 10–30% higher costs, weeks-long model adaptation and limited access for SMEs.

The market need is for trustworthy, low-barrier optimisation that can work on existing production lines with scarce data, variable inputs and legacy automation infrastructure. Operators also need transparent recommendations they can validate, rather than opaque black-box control systems that reduce trust and slow adoption.

AMOS combines data-efficient reinforcement learning, transfer learning and human-in-the-loop optimisation to adapt production parameters in real time from minimal data. PPO and Deep Q-Network agents learn from live sensor states and production outcomes, while policies learned on one material, machine or site are reused and fine-tuned for new batches or lines instead of retraining from scratch.

Operator corrections and validations become part of the learning signal, helping the system perform in early low-data deployments while explainable AI makes decisions auditable and supports trust. The platform is coupled with in-line IIoT quality control and multi-modal sensor fusion, so each measurement improves both process control and model learning.

AMOS will be deployed through an edge-first, microservices architecture designed for low latency, data sovereignty and retrofit integration with existing PLC/SCADA environments. The target outcomes include process stabilisation in around 4 hours instead of 48, at least 10% higher production efficiency, 15% less waste and energy use, 10–12% lower defect rates and 25% faster time-to-market.

The project will develop and integrate the AMOS platform modules: reinforcement-learning agents for live process optimisation, transfer-learning mechanisms across materials and production lines, a human-in-the-loop interface for operator feedback, and explainable AI functions for auditable decision-making.

The work will include integration of in-line IIoT quality-control sensors such as ultrasonic, capacitive and NIR spectrometry, together with multi-modal data fusion to feed viscosity, density, temperature and composition data into the optimisation loop. The team will implement an edge-first microservices architecture compatible with OPC-UA, MQTT and PLC/SCADA environments to enable low-latency, on-premises deployment on legacy lines.

The consortium will validate the system in industrial pilot conditions up to TRL 7, measuring performance against targets including faster process stabilisation, higher production efficiency, reduced waste and energy use, lower defect rates, faster time-to-market and improved operator trust.

• INEA d.o.o. — Technology provider (project coordinator)