New-generation AI-driven intelligent decision-making and inventory optimization in the full lifecycle of complex product manufacturing integrating LSTM and Q-learning

Smarter factories for everyday technology

From smartphones and laptops to cars and aircraft, many of the products people rely on every day depend on complex parts made through long, delicate manufacturing chains. These factories must decide how much to produce at each step, how much to keep in storage, and how to respond when demand suddenly jumps or falls. This study explores a new way to let artificial intelligence manage those decisions more intelligently, cutting waste and shortages while keeping high‑value products flowing.

Why making advanced products is so hard

Complex products such as semiconductor chips and aerospace components pass through many tightly linked stages, from cutting and bonding to testing and packaging. Demand for these products is often volatile: orders surge when a new device is popular and collapse when markets shift. Traditional planning methods struggle to cope with these swings. They either produce too much, tying up money in warehouses, or too little, causing painful stock‑outs where customers must wait or look elsewhere. At the same time, factories generate huge amounts of data from sensors, quality checks, and supply records that are difficult for human planners to digest in real time.

Combining two kinds of AI in a closed loop

The authors propose an integrated AI framework that links two complementary techniques: a forecasting module and a decision‑making module. The forecasting side uses a type of neural network called Long Short‑Term Memory (LSTM) to learn patterns in past orders and sensor signals, such as seasonal demand surges or slowdowns caused by equipment issues. Instead of treating this forecast as fixed, the system continuously refines it using new data. The decision side uses Q‑learning, a form of reinforcement learning in which a virtual decision‑maker experiments with different production choices in a simulated environment and learns which ones lead to lower costs and fewer shortages over time. Crucially, these two modules are connected in a feedback loop: forecasts guide decisions, and the outcomes of those decisions feed back to improve future forecasts.

Balancing cost, stock, and uncertainty

Real factories must juggle several cost types at once: the expense of making each item, the cost of holding extra stock on shelves, and the often much larger cost of running out and failing to meet orders. The framework models a production line with multiple stages and explicitly tracks inventory and backlogs at each one. It also adds a protective layer for uncertainty using a robust optimization technique that prepares for sudden shifts in demand rather than assuming past patterns will always hold. In practice, this means the system is less likely to be surprised by extreme scenarios, such as sudden market booms or disruptions in the supply chain, while still avoiding the habit of over‑producing “just in case.”

What the experiments show in real factories

To test the approach, the researchers applied it to real data from a semiconductor packaging and testing plant operating across five production stages. They compared their framework with traditional material requirements planning systems and several advanced AI‑based competitors. The LSTM forecasts reduced prediction error by about 40 percent compared with a classic statistical method, and the full framework cut total costs by 15.7 percent relative to the legacy system and by 8.3 percent compared with a strong modern benchmark. Remarkably, it achieved a zero stock‑out rate while turning inventory over 4.2 times per year, meaning parts did not sit idle for long. Further analyses showed that the system remained stable even when demand became more erratic, forecasts were less accurate, or sensor data became noisy—conditions that mirror everyday industrial reality.

Scaling up and spreading across industries

The study also examined how well the framework scales as factories grow or operate in other high‑value sectors. Simulations with more production stages showed only modest increases in computing time, suggesting that larger plants could adopt the approach without overwhelming their digital infrastructure. When adapted to aerospace component manufacturing, the method again reduced costs and kept shortages rare by simply tuning cost and capacity parameters to match that industry. The authors argue that, with support from edge computing hardware and careful system integration, the framework can be slotted into existing factory software used by many manufacturers today.

What this means for future production

In plain terms, the study demonstrates that giving factories a “digital co‑pilot” that both predicts demand and learns how to act on it can make production smoother, cheaper, and more reliable. By tightly linking forecasting with on‑the‑ground decisions and designing the system to cope with real‑world noise and surprises, the framework turns raw data into practical guidance for how much to produce and store at every step. For consumers, this kind of intelligence behind the scenes could mean fewer delays for new devices and more resilient supply of the advanced components that power modern life.

Citation: Jiang, Z., Dan, W. & Yu-fei, C. New-generation AI-driven intelligent decision-making and inventory optimization in the full lifecycle of complex product manufacturing integrating LSTM and Q-learning. Sci Rep 16, 11077 (2026). https://doi.org/10.1038/s41598-026-41629-6

Keywords: smart manufacturing, inventory optimization, semiconductor production, reinforcement learning, demand forecasting