A new study, led by Warren Jasper, professor at the US’ Wilson College of Textiles has demonstrated how machine learning can help reduce waste in textile manufacturing by improving the accuracy of colour prediction during the dyeing process.

The research, titled ‘A Controlled Study on Machine Learning Applications to Predict Dry Fabric Color from Wet Samples: Influences of Dye Concentration and Squeeze Pressure’, addresses one of the industry’s longstanding challenges: predicting what dyed fabric will look like once it dries.

Fabrics are typically dyed while wet, but their colours often change as they dry. This makes it difficult for manufacturers to determine the final appearance of the material during production. The issue is further complicated by the fact that colour changes from wet to dry are non-linear and vary across different shades, making it impossible to generalise data from one colour to another, according to the paper co-authored by Samuel Jasper.

“The fabric is dyed while wet, but the target shade is when its dry and wearable. That means that, if you have an error in coloration, you aren’t going to know until the fabric is dry. While you wait for that drying to happen, more fabric is being dyed the entire time. That leads to a lot of waste, because you just can’t catch the error until late in the process,” said Warren Jasper.

To address this, Jasper developed five machine learning models, including a neural network specifically designed to handle the non-linear relationship between wet and dry colour states. The models were trained on visual data from 763 fabric samples dyed in various colours. Jasper noted that each dyeing process took several hours, making data collection a time-intensive task.

All five machine learning models outperformed traditional, non-ML approaches in predicting final fabric colour, but the neural network proved to be the most accurate. It achieved a CIEDE2000 error as low as 0.01 and a median error of 0.7. In comparison, the other machine learning models showed error ranges from 1.1 to 1.6, while the baseline model recorded errors as high as 13.8.

The CIEDE2000 formula is a standard metric for measuring colour difference, and in the textile industry, values above 0.8 to 1.0 are generally considered unacceptable.

By enabling more accurate predictions of final fabric colour, the neural network could help manufacturers avoid costly dyeing mistakes and reduce material waste. Jasper expressed hope that similar machine learning tools would be adopted more widely across the textile sector to support efficiency and sustainability.

“We’re a bit behind the curve in textiles. The industry has started to move more toward machine learning models, but it’s been very slow. These types of models can offer powerful tools in cutting down on waste and improving productivity in continuous dyeing, which accounts for over 60 per cent of dyed fabrics,” stated Warren.

A study led by Warren Jasper shows machine learning can reduce textile dyeing waste by accurately predicting dry fabric colours from wet samples.
A neural network model trained on 763 samples achieved near-perfect accuracy, helping avoid costly errors.
Jasper urges wider adoption to boost sustainability and efficiency in continuous dyeing.

Fibre2Fashion News Desk (HU)

Birds flock in order to forage and move more efficiently. Fish school to avoid predators. And bees swarm to reproduce. Recent advances in artificial intelligence have sought to mimic these natural behaviors as a way to potentially improve search-and-rescue operations or to identify areas of wildfire spread over vast areas—largely through coordinated drone or robotic movements. However, developing a means to control and utilize this type of AI—or “swarm intelligence”—has proved challenging.

In a Proceedings of the National Academy of Sciences paper, an international team of scientists describes a framework designed to advance swarm intelligence—by controlling flocking and swarming in ways that are akin to what occurs in nature.

“One of the great challenges of designing robotic swarms is finding a decentralized control mechanism,” explains Matan Yah Ben Zion, an assistant professor at the Donders Center for Cognition at the Netherlands’ Radboud University and one of the authors of the paper.

“Fish, bees, and birds do this very well—they form magnificent structures and function without a singular leader or a directive. By contrast, synthetic swarms are nowhere near as agile—and controlling them for large-scale purposes is not yet possible.”

The research team, which included NYU scientists Mathias Casiulis and Stefano Martiniani, addressed these challenges by developing geometric design rules for the clustering of self-propelled particles. These rules are modeled using natural computation—similar to the “positive” or “negative” charges in protons and electrons that are foundational to the formation of matter.

Under these rules, active particles moving in response to external forces have an intrinsic property that causes them to curve—a quantity the researchers call “curvity.”

“This curvature drives the collective behavior of the swarm, which points to a means to potentially control whether the swarm flocks, flows, or clusters,” explains NYU’s Martiniani, an assistant professor of physics, chemistry, and mathematics.

Their conclusion was supported by a series of experiments in which the scientists showed that the curvature-based criterion controls robot-pair attraction and naturally extends to thousands of robots. Each robot was treated as having a positive or negative curvity, and similar to electric charge, this curvity controls the robots’ mutual interactions.

“This charge-like quantity, which we call ‘curvity,” can take positive or negative values and can be directly encoded into the mechanical structure of the robot,” explains Ben Zion.

“As with particle charges, the value of the curvity determines how robots become attracted to one another in order to cluster or deflect from one another in order to flock.”

Ben Zion, who, as an NYU student, previously developed microscopic swimmers, added, “Finding a design rule of geometric nature, such as curvature, makes it applicable to industrial or delivery robots or to cellular-sized microscopic robots that have the potential to improve drug delivery and other medical treatments.”

“The best part is that these rules are based on elementary mechanics, making their implementation in a physical robot straightforward,” adds Casiulis, a postdoctoral researcher at New York University’s Center for Soft Matter Research and NYU’s Simons Center for Computational Physical Chemistry.

“More broadly, this work transforms the challenge of controlling swarms into an exercise in materials science, offering a simple design rule to inform future swarm engineering.”

Nuclear energy is a leading option to power space exploration, but its success depends on reactors that can operate autonomously rather than relying on human operators in space.

To help make that vision a reality, Oak Ridge National Laboratory has built a non-nuclear test bed that mimics the conditions of a space nuclear reactor to overcome the high cost and strict regulations required for testing in a reactor environment. The research is published in the journal Energies.

This “hardware-in-the-loop” system—a system combining real hardware with computer models to simulate conditions—enables NASA and industry partners to rapidly develop and validate autonomous controls and hardware using cost-effective components and open-source software.

“Our test bed gives engineers the ability to push autonomous control systems to their limits in a safe, repeatable environment,” said ORNL’s Brandon Wilson. “That means we can identify and solve problems here on Earth—before astronauts rely on these systems millions of miles from home.”