Imagine buying three identical 3D printers. Despite being the same brand, the same model and even having similar serial numbers, each machine may behave slightly differently. Over time, and at scale, these differences can accumulate into significant manufacturing defects.
In order to address this issue, researchers from IMDEA Materials Institute, in collaboration with Lawrence Berkeley National Laboratory (USA), have developed an intelligent algorithm to detect these hidden ‘personalities’, optimising the reliability of automated manufacturing.
Published in Advanced Engineering Informatics, the system analyses subtle operational differences between theoretically identical machines and selects the most appropriate optimisation strategy.
In doing so, it reduces errors and improves the quality of the final product in parallel production systems, such as 3D printing farms.
One of the main challenges in large-scale automated manufacturing is that machines, even when they are the same make and model, never behave in exactly the same way.
These variations, or ‘noise’, in their performance can accumulate and lead to significant defects, reducing reproducibility and reliability, particularly in high-precision sectors such as additive manufacturing for architecture and the aerospace industry.
The new algorithm tackles this problem head-on.
Rather than applying a single solution to every machine, the system first performs a diagnostic assessment to create a unique performance profile for each one. Using statistical analysis, it quantifies the degree of variability between them. It then makes a key decision.
If the machines are sufficiently similar, it applies a joint optimisation strategy to maximise efficiency. If, on the other hand, it detects significant differences, it activates an individual optimisation strategy for each machine, prioritising accuracy.
To validate the method, the team carried out a study using three theoretically identical 3D printers. Although one would expect almost identical performance, the algorithm detected measurable differences between them and correctly identified that each required its own optimisation strategy.
The publication presents the results of both approaches, although the distributional analyses and divergence metrics clearly favoured single-device optimisation for the featured case study.
The pronounced separation observed in the density estimates of the printed pellets, together with the substantial pairwise divergence values, indicates that each printer operated within a distinct output regime.
Consequently, the most appropriate approach for this case study was to optimise each device individually, ensuring that the specific characteristics of each printer were taken into account.
According to the researchers involved in the study, “the results demonstrated significantly faster convergence and a substantial reduction in errors in the weight of printed parts compared with treating all machines equally and thus failing to correct appropriately for individual biases.”
“Even mass-produced machines may have their own operational ‘personality’. Our system learns these differences and uses them to our advantage, determining whether it is more efficient to treat them as a team or as individuals.”
“This not only improves accuracy, but also saves resources by avoiding failed experiments, a key step towards the fully automated laboratories and factories of the future,” the researchers concluded.
Although the study focused on 3D printing, its methodology is applicable to other fields that rely on high-throughput experimentation, such as the discovery of new materials, chemical synthesis and sensor calibration.
The study was carried out by Dr. Christina Schenk, Miguel Hernández del Valle, Luis Calero and Dr. Maciej Haranczyk from IMDEA Materials Institute, together with Dr. Marcus Noack from Lawrence Berkeley National Laboratory.
This work was supported by the MAD2D-CM project on Two-Dimensional Disruptive Materials funded by the Community of Madrid, the Recovery, Transformation and Resilience Plan, Spain, and NextGenerationEU from the European Union. MN’s contribution to this work was supported by the Center for Advanced Mathematics for Energy Research Applications (CAMERA), funded jointly by the Advanced Scientific Computing Research (ASCR), United States and Basic Energy Sciences (BES) programs in the U.S. Department of Energy’s Office of Science, under Contract No. DE-AC02-05CH11231. Additionally, CS acknowledges financial support from the Spanish Ministry of Science and Innovation through a Ramón Cajal grant (Grant No. RYC2024-048744-I), financed by MICIU/AEI/10.13039/501100011033 and FSE+.