Machine Learning Cracks the Optical Code of Lithium Bromide Solutions

In the world of absorption chillers and heat pumps, a humble salt solution does the heavy lifting. Water and lithium bromide form the working fluid at the heart of countless industrial cooling systems, quietly absorbing vapor and releasing heat as they cycle through machinery. Now, a team of researchers has shown that a carefully tuned machine learning model can predict one of this fluid’s most diagnostically important properties, its refractive index, with a precision that rivals the laboratory instruments used to measure it. The work, published in Results in Chemistry, also delivers a surprise: a trace of surfactant, long assumed to be optically irrelevant, measurably bends light in these solutions.

The refractive index, a measure of how much light slows and bends as it passes through a material, is far more than an optical curiosity in this context. Because the index shifts systematically with the concentration of dissolved lithium bromide, engineers can use optical techniques such as interferometry and inline refractometry to monitor solution strength without ever touching the fluid. That matters enormously in absorption refrigeration, where the concentration of the lithium bromide solution must be tightly controlled. Let the concentration drift too high and crystals form, clogging pumps and heat exchangers in one of the most feared failure modes of the technology. Accurate, real-time optical sensing is therefore a safety and efficiency tool, not merely a laboratory convenience.

Until recently, the optical data available for lithium bromide solutions were sparse. Most historical measurements were taken at a single wavelength, the sodium D-line at 589.3 nanometers, over narrow ranges of temperature and composition, and the resulting datasets were often inconsistent. A major change came when Pérez de Luco and colleagues published a comprehensive experimental database spanning multiple wavelengths, a wide sweep of lithium bromide mass fractions, and temperatures from 23 to 47 degrees Celsius, with and without 150 parts per million of 1-octanol, a surfactant commonly added to enhance heat and mass transfer. The new study builds directly on that dataset, which contains 3,185 individual refractive index measurements, to construct predictive models that capture all of these variables simultaneously.

The researchers benchmarked five machine learning architectures: CatBoost, a gradient-boosting method; a multilayer perceptron neural network; a heterogeneous ensemble combining support vector machines, decision trees, and k-nearest-neighbor regressors; AdaBoost; and Random Forest. Rather than hand-tuning their models, the team employed Coupled Simulated Annealing, an optimization strategy that runs multiple search trajectories in parallel under a shared acceptance rule. This cooperative scheme helps the search escape local minima that trap conventional single-path optimizers. The tuning was performed strictly on a validation subset carved out of the training data, with an independent test set of 637 points held back until the very end, a discipline designed to prevent the subtle data leakage that can inflate reported accuracy.

The verdict was emphatic. CatBoost emerged as the clear winner, achieving a coefficient of determination of 0.99988 on the test set, a root mean square error of 0.00069, and an average absolute relative error of just 0.0366 percent. That error is within roughly one to three times the stated experimental uncertainty of the underlying measurements, which stands at plus or minus 0.0003 in refractive index. The neural network and the heterogeneous ensemble followed closely, while AdaBoost and Random Forest trailed with errors roughly double those of CatBoost. Notably, the gap between training and testing performance for CatBoost was a mere 1.47 percent in RMSE, strong evidence that the model genuinely learned the underlying physics rather than memorizing its training examples.

Perhaps the most scientifically interesting result came from a Monte Carlo sensitivity analysis, which quantified how strongly each input variable drives the refractive index. Wavelength dominated with a sensitivity index of 6.97, consistent with the well-known phenomenon of optical dispersion. Temperature ranked second at 4.95, reflecting its dual influence on solution density and molecular polarizability. But in third place, with an index of 3.67, sat surfactant presence, ahead of lithium bromide mass fraction itself at 1.52. Comparing matched experimental pairs, the team found that adding 150 ppm of 1-octanol consistently lowers the refractive index by about 0.0004 to 0.0012 units, a small but systematic effect that earlier optical models of absorption systems had universally ignored.

The practical consequences of that neglected effect are striking. According to the authors’ analysis, an optical diagnostic calibrated on surfactant-free solutions but deployed on a surfactant-containing system would misread the lithium bromide concentration by roughly 1.5 to 2.5 weight percent. In an absorption chiller, where crystallization prevention depends on knowing concentration to within about half a percent, that is an error no operator can afford. The new models are, to the authors’ knowledge, the first predictive tools to explicitly and accurately quantify this surfactant-induced optical shift across the full experimental range, allowing engineers to interpret optical signals correctly even when heat-transfer additives are present.

The team also stress-tested the winning model with a grouped leave-one-out validation scheme, deliberately withholding entire physical regimes rather than random data points. When all measurements at 655 nanometers were excluded, the model had to extrapolate to an unseen wavelength, and its RMSE rose from 0.00069 to 0.00124. Excluding the high-temperature band of 44 to 47 degrees Celsius and the high-concentration region above 0.55 mass fraction produced similar, slightly smaller degradations. Even under these demanding conditions, the coefficient of determination never fell below 0.9979, and the translated concentration error of about 0.39 percent remains within the tolerances of typical chiller monitoring. Intriguingly, the wavelength-exclusion test produced the largest error, mirroring the Monte Carlo finding that dispersion is the dominant variable, an agreement between two independent analyses that strengthens confidence in the physical interpretation.

Speed matters as much as accuracy for real-world deployment, and here the models also shine. Once trained, CatBoost delivers predictions in milliseconds with a small memory footprint, making it suitable for embedding in optical sensors, computational fluid dynamics codes, and digital twin platforms for absorption chillers. The authors envision the model serving as a surrogate that replaces expensive property calculations inside cycle simulators and CFD solvers, or feeding real-time control strategies that keep systems safely away from the crystallization boundary. Designers of lab-on-a-chip devices and liquid-core optical fibers that might contact lithium brine fluids could likewise use the models to predict optical behavior without extensive characterization campaigns.

The study is candid about its limits. The models are data-driven and valid only within the training domain of 23 to 47 degrees Celsius, wavelengths from 436.1 to 655 nanometers, and mass fractions up to 0.6199 kilograms per kilogram; extrapolation beyond these bounds is not recommended. The binary surfactant variable also captures only the fixed 150 ppm dosage used in the source experiments, and the sensitivity ranking for a discrete on-off variable must be interpreted differently from that of continuous inputs. Still, the authors argue that the framework itself, from Coupled Simulated Annealing optimization to grouped cross-validation, is fully generalizable, and future work could embed physical constraints such as molar refractivity and thermodynamic consistency directly into the training process. For now, the message is clear: with the right algorithm and rigorous validation, machine learning can turn a sprawling optical dataset into a fast, trustworthy instrument in its own right, one that sees both the salt and the subtle fingerprint of the surfactant dissolved within it.

Subject of Research: Machine learning prediction of the refractive index of surfactant-containing aqueous lithium bromide solutions

Article Title: Accurate refractive index modeling of surfactant-containing LiBr–water solutions via optimized machine learning techniques

Article References: Al Bustanji, S., Mostafa, S. A., Prajapati, A. A., Gowrishankar, J., Shakir, A. K., Bharti, R., Ernazarov, M., Madaminov, B., & Hekmatyar, Z. (2026). Accurate refractive index modeling of surfactant-containing LiBr–water solutions via optimized machine learning techniques. Results in Chemistry, 31, Article 103874. https://doi.org/10.1016/j.rechem.2026.103874

Image Credits: AI Generated

DOI: 10.1016/j.rechem.2026.103874

Keywords: refractive index, lithium bromide, machine learning, CatBoost, surfactant, absorption chillers, Coupled Simulated Annealing, interferometry, sensitivity analysis, Monte Carlo simulation, optical sensing, hyperparameter optimization