Chemists have long known that the carbon–fluorine bond is one of the toughest single bonds in organic chemistry, with dissociation energies of roughly 485 to 543 kilojoules per mole. That strength is what makes per- and polyfluoroalkyl substances, or PFAS, so durable in the environment, and it is also what makes destroying them so difficult. A new computational study now offers a systematic map of what it takes to break that bond in tetrafluoromethane, the simplest and most persistent perfluorocarbon, and pairs that map with a machine learning model that can predict activation energies without running a full transition-state search for every candidate catalyst.
The work, published in Frontiers in Chemistry, was carried out by a team led by Sha Li, Yuan Liu, Bing Yu and Lingbo Qu, with collaborators at the Henan Academy of Sciences, Shantou University, Sinochem Dongda and Zhengzhou University. The researchers used density functional theory calculations to examine how tetrafluoromethane, commonly written CF4, dissociates on three distinct classes of catalyst surfaces: extended transition metal surfaces, single-atom alloys in which one metal atom sits inside a copper lattice, and single-atom catalysts anchored to a silicon carbide support. In every case they focused on the first, decisive step: the direct cleavage of one C–F bond to leave adsorbed CF3 and fluorine fragments on the surface.
CF4 is not an arbitrary test case. It has a highly symmetric tetrahedral structure, a C–F bond dissociation energy of about 543 kilojoules per mole, an atmospheric lifetime exceeding 50,000 years, and a place among the most potent greenhouse gases. The semiconductor manufacturing and aluminum electrolysis industries are its primary anthropogenic sources, and it is also frequently generated as a terminal degradation product when longer-chain PFAS are incompletely decomposed by heat or chemistry. Any strategy that aims to truly destroy PFAS, rather than merely transfer it from water into a secondary waste stream, must eventually confront this molecule.
On the nine transition metal surfaces studied, which included iron, molybdenum, nickel, copper, rhodium, platinum, gold, ruthenium and rhenium, the team found that intact CF4 interacts only weakly with the surface, with binding energies between roughly −0.02 and 0 electron volts. Once the bond begins to break, however, a clear pattern emerges. The activation energy for C–F cleavage scales linearly with the reaction energy, following the equation Ea = 0.56 × ΔE + 2.17 electron volts, and it also scales linearly with the fluorine binding energy, following Ea = 0.84 × BE(F) + 4.44 electron volts. These are Brønsted-Evans-Polanyi type correlations, the same kind of scaling relations that have long guided catalyst design for hydrogenation and other reactions.
The comparison with the well-known C–H analogue is striking. For methane dissociation on transition metal surfaces, the established relation is Ea = 0.68 × ΔE + 1.04 electron volts. For C–F cleavage, the slope is lower, at 0.56, meaning the barrier is less sensitive to changes in reaction energy, but the intercept is dramatically higher, at 2.17 electron volts. Quantitatively, the C–F cleavage barrier runs 0.5 to 1.5 electron volts higher than the corresponding C–H barrier across the surfaces examined. The authors attribute this penalty to the extreme thermodynamic stability of the C–F bond, strong ionic character in the polarized Cδ+–Fδ− unit, and pronounced Pauli repulsion as the fluorine lone pairs overlap with surface electron density in the transition state.
When single metal atoms were embedded into a Cu(111) surface to form single-atom alloys, the picture changed. The relationship between activation energy and reaction energy became volcano-shaped rather than monotonic, describable by two linear branches that meet near a reaction energy of −0.53 electron volts. Yet the correlation between activation energy and fluorine binding energy remained firmly linear, with Ea = BE(F) + 5.06 electron volts and a coefficient of determination of 0.98 across the nine dopant metals tested. Extending the calculations to silver- and gold-hosted alloys confirmed that a common linear trend holds across all three host families, with a combined fit of Ea = 0.90 × BE(F) + 4.71 electron volts. Within the binding energy range covered, the two fits differ by no more than 0.15 electron volts.
The silicon carbide-supported single-atom catalysts broke the pattern entirely. Here each metal atom sits in a hollow site coordinated to three surface carbon atoms and three subsurface silicon atoms, a much lower coordination number than the nine-fold environment in the copper-based alloys. Both the activation energy versus reaction energy and the activation energy versus fluorine binding energy became volcano-shaped. For nickel, platinum and rhodium, whose fluorine binding energies are comparable to their alloy counterparts, the C–F activation energies were nonetheless lower by 0.4 to 0.8 electron volts. Electron localization function and Bader charge analyses showed that the silicon carbide support induces orbital-selective splitting of the metal d states and produces an earlier transition state that retains residual polar-covalent C–F bonding alongside ionic metal–fluorine stabilization and support polarization.
The machine learning component targeted the single-atom alloys, where the linear Ea–BE(F) relation makes fluorine binding energy a reliable proxy for the activation barrier. The team built a dataset of density of states and fluorine binding energies for 90 copper-, silver- and gold-based SAA systems, then trained three architectures: a convolutional neural network, a long short-term memory network and a Transformer. The CNN won decisively, achieving a test mean absolute error of 0.163 electron volts and a coefficient of determination of 0.949, while the LSTM and Transformer reached errors of 0.338 and 0.403 electron volts respectively. The best CNN variant, which encodes energy-axis information directly into every channel of the spectral input, improved further to an ensemble test error of 0.144 electron volts with an R² of 0.968.
That accuracy is meaningful in context. Density functional theory calculations themselves carry errors on the order of 0.1 to 0.2 electron volts, so a surrogate model that predicts binding energies to within 0.144 electron volts approaches the reproducibility limit of the underlying method. Across the 16-sample test set, 81.2 percent of predictions fell within 0.2 electron volts of the calculated values and none exceeded 0.3 electron volts. Combining these predicted binding energies with the host-inclusive scaling relation allows researchers to estimate C–F activation barriers for alloy candidates from electronic structure alone, skipping the computationally expensive transition-state searches that would otherwise be needed for each new composition.
The authors are careful to note the boundaries of what they have shown. All energetics were computed at the PBE level without zero-point or finite-temperature corrections, spin-state effects were not explicitly considered, and the scaling relations describe electronic-energy trends rather than free-energy relationships. The turning points in the volcano curves are empirical crossovers for the specific catalyst sets studied, not universal boundaries. The machine learning workflow, while demonstrated here for single-atom alloys, would need class-specific training data to extend to other catalyst families. Even so, the study delivers something the field has lacked: a unified, quantitative account of why C–F bond activation is so hard, which descriptors govern it on each surface type, and how a data-driven shortcut can now screen candidate catalysts for tetrafluoromethane abatement before a single experiment is run.
Subject of Research: Computational prediction of C–F bond activation energies in tetrafluoromethane using scaling relations and machine learning across transition metal, single-atom alloy and silicon carbide-supported catalysts
Article Title: Combined scaling relations and machine learning for predicting C-F bond activation energies in tetrafluoromethane
Article References: Li, S., Xu, W., Qin, C., Chen, H., Gao, X., Liu, Y., Yu, B., & Qu, L. (2026). Combined scaling relations and machine learning for predicting C-F bond activation energies in tetrafluoromethane. Frontiers in Chemistry, 14, Article 1984156. https://doi.org/10.3389/fchem.2026.1984156
Image Credits: AI Generated
DOI: 10.3389/fchem.2026.1984156
Keywords: PFAS, tetrafluoromethane, C–F bond activation, density functional theory, scaling relations, machine learning, convolutional neural network, single-atom alloys, single-atom catalysts, fluorine binding energy, catalyst screening, green chemistry
News Source: Blake Davidson. (November 6, 2026). Machine learning maps the hardest bond in the fight against forever chemicals. Scienmag.
