The team, led by Ananth Govind Rajan, Associate Professor in the Department of Chemical Engineering at IISc, constructed a data-driven computational framework that maps nearly 10,000 elementary chemical reactions involved in converting CO2 into fuels and chemicals over a copper catalyst. The study, published in Nature Communications, demonstrates that a small, hand-picked reaction network can lead a model spectacularly astray, while a comprehensive, machine-learning-accelerated network brings predictions into line with laboratory reality. The work offers a template that could extend well beyond CO2 hydrogenation to other catalytic processes central to a sustainable energy economy.

The scale of the problem is easy to underestimate. In CO2 hydrogenation, CO2 reacts with hydrogen over a catalyst to yield products such as methanol and carbon monoxide, but each product emerges from a tangled web of intermediates adsorbed on the metal surface. Traditional mechanistic modelling relies on chemical intuition to select perhaps a few dozen likely reactions, because modeling even one reaction with methods such as density functional theory demands enormous computational resources. That intuition, the IISc researchers argued, carries a hidden risk: the network might quietly omit the one step that actually controls the outcome.

First author Anand Mohan Verma, who carried out the work as a CV Raman Postdoctoral Fellow at IISc and is currently an Assistant Professor at Motilal Nehru National Institute of Technology Allahabad, framed the central dilemma of the field. Mechanistic modellers can never be certain that a curated reaction network has not excluded the critical step, and that uncertainty undermines confidence in every downstream prediction. The new framework was designed specifically to eliminate that blind spot by replacing intuition with systematic enumeration.

The researchers began by generating a carefully curated database of 152 reactions using quantum-mechanical simulations, establishing a foundation of reliable energetics. They then trained machine learning models to rapidly predict activation energy barriers for additional reactions, borrowing the pattern-recognition strengths of modern artificial intelligence to sidestep the expense of calculating each barrier from first principles. In parallel, they deployed automated enumeration tools to identify every possible single-step reaction involving 105 distinct surface species. Combining these approaches expanded the original seed network to 9,389 elementary reactions, a catalog several orders of magnitude larger than anything a conventional modelling effort would attempt.

The payoff was immediate and dramatic. When the team modeled CO2 hydrogenation using only the initial 152 reactions, the network predicted that formic acid, not methanol, would emerge as the dominant product, and it badly underestimated how much CO2 would be converted overall. Only after expanding the network to include the thousands of previously overlooked reactions did the predictions align with experimental observations from the researchers’ own work and from published studies. Rajan emphasized that this discrepancy shows how small networks can produce qualitatively wrong answers while appearing perfectly rigorous on their own terms.

Plugged into a kinetic model, the expanded network predicted an approximately 40-fold increase in CO2 conversion compared with the sparse network, and it correctly identified methanol and carbon monoxide as the major products, matching what experiments actually deliver. Experimental validation was performed by collaborators G Valavarasu and Santhosh Kotni at Hindustan Petroleum Corporation Limited’s Green Research and Development Centre, and by Amol Amrute and colleagues at the Agency for Science, Technology, and Research in Singapore. The machine learning components drew on expertise from Ambedkar Dukkipati, Professor in the Department of Computer Science and Automation at IISc, underscoring the interdisciplinary character of the project.

The most conceptually striking result concerned hydrogen transfer. The enlarged network revealed that in several key steps, hydrogen can be delivered to reaction intermediates directly as intact molecular H2, rather than only through individual hydrogen atoms dissociated on the surface. Quantum-mechanical calculations performed explicitly on those steps confirmed that this molecular pathway can be particularly favourable for hydrogen transfer to oxygen-containing intermediates. This finding challenges a foundational assumption taught in catalysis courses, namely that H2 must first split into atoms before it can participate in surface chemistry, and it emerged purely because the network was large enough for the pathway to appear.

Shivam Chaturvedi, a PhD student in the Department of Chemical Engineering at IISc and co-author of the study, noted that the molecular hydrogen transfer pathway surfaced only because the reaction network was sufficiently expansive, and that the observation held up when the team returned to compute those specific steps with full quantum-mechanical rigor. The finding also carries practical implications for catalyst design. Catalysts that interact more strongly with H2 could potentially enhance the pathways leading to methanol, offering a concrete design principle for engineers seeking to improve the yield of this valuable fuel and chemical feedstock from CO2.

Beyond its immediate insights into copper-catalyzed CO2 hydrogenation, the framework represents a general strategy for exploring complex catalytic chemistry. By combining quantum-mechanical simulations for ground-truth energetics, machine learning for rapid interpolation across vast reaction spaces, automated enumeration for exhaustive coverage, and kinetic modelling for connection to observable outcomes, the approach can in principle be applied to other industrially important processes, including CO2 reduction on alternative catalysts, nitrogen reduction for ammonia synthesis, and water splitting for green hydrogen production. As decarbonization efforts intensify worldwide, tools that can map reaction landscapes comprehensively, rather than selectively, may prove decisive in turning atmospheric carbon into useful chemistry at scale.

Subject of Research: Data-driven computational mapping of catalytic CO2 hydrogenation reaction networks on copper

Article Title: Using machine learning and computation to unveil mechanisms for CO2-to-fuel conversion

Article References: Using machine learning and computation to unveil mechanisms for CO2-to-fuel conversion. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: CO2 hydrogenation, machine learning, copper catalyst, methanol synthesis, reaction networks, computational catalysis, density functional theory, kinetic modelling, catalyst design, hydrogen transfer, Nature Communications, carbon recycling

The research team, led by Evangelia Pasidi, Nikolaos Stratis and Patroklos Vareltzis, prepared sunflower-oil emulsions fortified with five different concentrations of vitamin D3, ranging from 10 to 100 micrograms per milliliter. The droplets were homogenized by ultrasound into a consistent coarse emulsion with a mean droplet diameter of roughly 2.3 to 2.5 micrometers, a size deliberately held constant across batches so that any differences in digestion behavior could be attributed to vitamin loading rather than to physical structure. The emulsions were then run through the INFOGEST protocol, the internationally standardized in vitro digestion method that mimics the oral, gastric and intestinal phases with realistic enzymes, bile salts, pH values and 37-degree incubation.

The critical action takes place in the intestinal phase, where pancreatin and bile salts are added and the pH is raised to 7. The researchers sampled independent digestion tubes every five minutes for two hours, isolated the micellar fraction by high-speed centrifugation and filtration, saponified the samples to liberate the vitamin from its lipid carriers, and quantified cholecalciferol by high-performance liquid chromatography with ultraviolet detection. The resulting time courses revealed a striking pattern: micellar vitamin D3 climbed rapidly during the first thirty minutes of intestinal digestion and then settled onto a plateau that persisted for the remainder of the two-hour experiment, with only small, statistically insignificant fluctuations.

Speed, it turns out, depends on dose. During the rising phase, micellar vitamin concentration grew at 0.093 micrograms per milliliter per minute for the most dilute emulsion, 0.323 for the intermediate one, and 1.112 for the most concentrated, differences that were statistically significant. But when the data were normalized to each emulsion’s initial vitamin load, the rates converged to statistically indistinguishable values between 0.006 and 0.01 per minute. In other words, the shape of the incorporation curve is an intrinsic property of the digestion system, not of how much vitamin was poured in at the start.

Concentration did matter, however, for efficiency. Within just five minutes of the intestinal phase, the dilute 10 microgram-per-milliliter emulsion had already delivered 60 percent of its initial vitamin into the micellar phase, while the richer formulations managed only 36 and 32 percent. By the plateau, the lowest dose retained a bioaccessibility index of 0.69, the highest of the series, whereas the intermediate concentrations plateaued near 0.53 to 0.56 and the highest dose recovered to 0.61. The authors attribute this inverse relationship to a finite pool of mixed micelles: bile salts are added once at a fixed concentration in the static protocol, fatty acid release is similar across samples because the oil-to-water ratio is constant, and so the micellar carrying capacity is capped. At low vitamin loads, nearly every encounter between a micelle and a dissolved vitamin molecule succeeds; at high loads, the micelles saturate and excess vitamin is left behind, possibly forming self-aggregates that never make it into the absorptive fraction.

To describe these dynamics quantitatively, the team built two competing mechanistic models, each partitioning the vitamin into an unreleased pool trapped in the lipid phase, a released pool in the bulk aqueous phase, and the micellar destination. Model 1 uses first-order kinetics throughout, with the release and micellization rates multiplied by a lipolysis-progress term, the running ratio of free fatty acids produced to the maximum producible, plus a loss term accounting for precipitation or degradation that prevents bioaccessibility from ever reaching 100 percent. Gastric lipolysis was measured separately by titration and proved minimal, reaching only 1.68 percent hydrolysis, a result consistent with the known ability of the Tween 80 emulsifier to displace gastric lipase from droplet surfaces.

Model 2 goes further, introducing an explicit population of micelles with a maximum number constrained by bile salt availability, a vitamin formation flux that slows as micellar capacity fills, and a dynamic loading term describing how much vitamin each individual micelle can carry. That loading follows Michaelis-Menten-style saturation in the released vitamin concentration and includes a swelling factor, reflecting prior evidence that solubilizing hydrophobic molecules physically enlarges micelles. Parameters were estimated with an Enhanced Scatter Search optimizer in MATLAB, and both models were rigorously validated using leave-one-out cross validation across the five concentrations, a fitting discipline that tests whether the equations genuinely generalize rather than merely memorize the training data.

Model 2 emerged as the stronger performer. Its mean absolute percentage error on training data was 12.3 percent versus 17.7 percent for Model 1, a statistically significant improvement, and its test-set error of 16.9 percent came with nearly 30 percent less variability than its rival. Perhaps most tellingly, both models independently converged on a lipolysis rate constant of 0.04 per minute with essentially zero variance across folds, indicating that the fat-hydrolysis machinery of the system was captured with high confidence. The micellization constants were consistently higher than every other rate constant in both models, supporting a key physiological conclusion: micellar incorporation is rapid and is not the rate-limiting step in making vitamin D3 bioaccessible. Instead, the bottleneck lies upstream, in the pace at which lipolysis frees fatty acids to build the micellar ferry in the first place.

The authors are candid about the model’s soft spots. Several parameters, notably the maximum micelle number and the intrinsic per-micelle loading capacity in Model 2, are structurally meaningful but practically unidentifiable from the current data, because they enter the equations as a product whose factors trade off against each other. They propose that future work directly measure micelle counts during digestion and track whole-digesta vitamin concentrations to close the mass balance, which would tighten these estimates. The study is also limited to a single coarse sunflower-oil emulsion under static conditions, whereas real digestion involves dynamic pH regulation and gallbladder bile delivered gradually rather than as a single bolus.

Even so, the implications for food and supplement design are considerable. Because bioaccessibility peaks at low doses and saturates beyond roughly 25 micrograms per milliliter, simply fortifying foods with more vitamin D3 yields diminishing returns once micellar capacity is exhausted; smarter strategies such as nanoemulsion carriers, which previous work has shown to boost D3 bioaccessibility relative to coarse emulsions, may achieve more. The integrated ODE framework itself is arguably the study’s most transferable product: by coupling release, lipolysis and micellization into one predictive engine, it offers supplement manufacturers and fortified-food formulators a computational tool for engineering release profiles in silico before committing to bench trials, and a foundation for embedding digestion chemistry into physiologically based models of nutrient delivery.

Subject of Research: Kinetic modelling of vitamin D3 bioaccessibility during simulated intestinal digestion

Article Title: Vitamin D 3 concentration behavior during intestinal in vitro digestion: an experimental and kinetic modelling approach

Article References: Vitamin D 3 concentration behavior during intestinal in vitro digestion: an experimental and kinetic modelling approach. (n.d.). https://doi.org/10.1016/j.jafr.2026.103310

Image Credits: AI Generated

DOI: 10.1016/j.jafr.2026.103310

Keywords: vitamin D3, bioaccessibility, in vitro digestion, INFOGEST, mixed micelles, lipolysis, kinetic modelling, ordinary differential equations, emulsions, HPLC, nutrient delivery, food fortification