“Agri-food waste” is defined as inedible or wasteful plant- and animal-based products such as fruit and vegetable peels, spent grains, shells, and post-consumer food wastes. The organic components of the various AFW streams reported in Table 1 consist largely of lignocellulosic fractions, including cellulose (11-65 wt.%), hemicellulose (6-50 wt.%), and lignin (3-36 wt.%), with each macromolecule determining its specific mechanism for creating carbon nanostructures. The chemical composition of AFW affects CQD formation and properties. Carbohydrates, in particular, are considered superior carbon sources for carbon quantum dot (CQD) synthesis due to their chemical structure, high carbon content, and distinct thermal decomposition behavior. Natural polymers present in AFW, such as cellulose, hemicellulose, and lignin, provide a rich carbon framework that readily converts into nanostructured carbon materials under thermal treatment^[2,7]. For example, rice straw and nut shells are commonly used agricultural byproducts, which provide the main carbon backbone by the dehydration and polymerization of cellulose and hemicellulose^[10,11]. On the other hand, the proteins and lipids in food waste are an inherent source of nitrogen, providing in situ self-doping that results in improved surface reactivity and polarity of CDs, avoiding the use of high-cost external dopants^[12]. Residual lipids play an important role in modulating the surface hydrophobic-hydrophilic balance. This is crucial for preventing particles from agglomerating in order to guarantee ultra-long-term colloidal stability when they are dispersed in different thermal base fluids ^[7]. Finally, the combined composition enables the engineering of CDs with a desired surface chemistry based on the molecular structure of the starting material. The typical physicochemical characteristics of various AFW precursors providing a rich carbon framework are summarized in Table 1.

The key to developing CDs for high-performance thermal fluids lies in understanding their basic structural classifications and inherent physics. CDs fall broadly within the general class of zero-dimensional (0D) carbonaceous nanomaterials. However, they do not represent a monolithic family of materials. Rather, according to their carbonization level within the individual structure and precursor source, they can be classified into three main structural classes for taxonomic purposes: (i) graphene dots (GDs), composed of one or few layer sheets of graphene with clearly defined sp² crystalline domains; (ii) carbon dots (CDs), which are spherical nanoparticles with a graphitic core that is dislocated yet linked locally with localized quantum confinement; and (iii) carbonized polymer dots (CPDs), composed of deep cross-linked amorphous polymeric chains without well-defined structured crystalline networks^[12,24,25].

Morphologically, HTC-derived CDs prepared from biomass through subcritical liquid transformation follow a core-shell structure: a core consisting of carbonized sp²-hybridized clusters responsible for electron flow and primary density, surrounded by a shell that is amorphous, defective, and filled with functional groups derived from the original food waste^[18,19]. When employing these structures for thermal fluid applications, the most important design consideration is to maintain an optimum balance between the degree of graphitization in the core (responsible for intrinsic heat conduction) and the passivating effect in the shell (responsible for fluid colloidal stability). HTC of heterogeneous AFW leads to CPDs since cross-linking polymerization prevails during the subcritical aqueous reaction pathway. This review focuses largely on CPD-type CDs, unless otherwise noted, although changing synthesis parameters allows tuning the product distribution toward more graphitized CD-type products.

The selection of the appropriate thermochemical process is key to determining the ultimate shape and surface properties of carbonaceous products. Dry torrefaction involves conventional heating at temperatures of 200-300 °C in an inert atmosphere, aimed at eliminating moisture and volatiles from the feedstock, which results in high-calorific solid biofuels (torrefied biomass)^[10]. This solid-state transformation produces bulk, non-dispersible, hydrophobic macro-carbon structures, which are not suitable for stable colloidal suspensions.

In contrast, HTC takes place in a subcritical water medium and enables liquid-mediated reaction kinetics favorable for CD production. The HTC temperature (180-250 °C) range is similar to that of dry torrefaction; however, it occurs in an aqueous subcritical system that enables liquid-mediated nucleation instead of solid-state carbonization. In this aqueous, high-pressure medium, AFW macromolecules are rapidly hydrolyzed into reactive monomeric units. The monomers can then undergo polymerization and nucleation into 0D nanostructures^[14]. When compared to dry torrefaction, HTC provides a more energy-efficient alternative with improved control over particle size and surface functionalization, thus providing the preferred approach for advanced nanofluid applications^[6]. A detailed comparison between these strategies and their implications for CD synthesis is presented in Table 2.

HTC-based synthesis of CDs involves a complex liquid-based route that differs significantly from conventional solid-state thermochemical processes. At subcritical water temperatures (180-250 °C), autogenous pressure is produced in the reactor. Formation of CDs involves four stages [Figure 2]: (i) hydrolysis of AFW macromolecules (carbohydrates, proteins, and lipids) into reactive monomers (e.g., glucose and amino acids); (ii) dehydration and fragmentation to form soluble furan-like intermediates; (iii) polymerization including cross-linking into polymer clusters that often proceed via Maillard-type interactions; and (iv) aromatization and nucleation leading to the formation of a stable graphitic core surrounded by a functionalized shell once critical supersaturation is reached^[6,7,14]. Superior performance of AFW-based CDs as thermal nanofluid additives arises primarily due to their unique physicochemical properties. In terms of morphology, CDs appear as quasi-spherical structures with ultra-small diameters (usually less than 10 nm), resulting in a high surface-to-volume ratio for better energy transport ^[14].

A distinctive feature is their “self-passivated” surface chemistry derived from the nitrogenous and oxygenated species present in AFW. In addition to the good thermal conductivity, these nanomaterials have numerous hydroxyl (-OH), carboxyl (-COOH), and amino (-NH₂) groups on their surface, which render them extremely hydrophilic and highly dispersible in water^[15]. The particles are more stable than traditional metallic nanoparticles due to the high density of surface groups; this results in excellent colloidal stability compared to other metallic nanoparticles, which often suffer from serious van der Waals forces and hence experience agglomeration and quick sedimentation in aqueous solutions under gravity^[26].

A high absolute zeta potential value (typically greater than |30| mV) enables the generation of strong electrostatic repulsion and steric hindrance between nanoparticles^[26]. The nanoparticles do not undergo agglomeration and sedimentation processes. There is no need for the use of toxic or foaming surfactants^[27]. This is an important feature of these nanoparticles in terms of engineering requirements for preparing high-performance thermal nanofluids because they do not exhibit any inconsistency in their thermophysical properties and do not pose the possibility of clogging in micro-channels in the case of the application of the fluids in continuous flow devices^[26].

The chemical synthesis of CDs from heterogeneous AFW using HTC involves complex non-linear reaction kinetics with multiple input variables acting on each other at the same time (e.g., temperature, residence time, biomass-to-water ratio, precursor composition). The parameters are typically varied individually through traditional OFAT screening, neglecting the simultaneous interaction between them. Such interactions are the most important factors controlling the final product quality^[25]. Uncontrolled nanoparticle formation results in inconsistent core graphitization and varying surface passivation, hampering efforts to formulate consistent thermal nanofluids. The field needs to transition from trial-and-error to a data-driven approach. Several recent examples describe the use of computational analytics, statistical modeling, and machine learning (ML) to predictably engineer the interaction between precursors, processes, and properties within a limited parameter space^[12,28].

However, extending the approach to larger untried precursor/process combinations has not yet been done, although it is a logical next step. The statistical DoE offers a middle ground before implementing ML models. Using the Taguchi orthogonal array method, the full factorial design (> 100 runs) can be reduced to a much smaller orthogonal design (< 20 runs). The approach eables identifying key HTC control factors such as temperature, residence time, and biomass-to-water ratio without excessive resources^[8]. RSM approaches, namely the Box-Behnken and central composite designs, can generate response surfaces to map the relationship between precursor and process parameters to optimize the yield of CDs, their size (d < 10 nm), and the fluorescence quantum yield^[10].

DoE creates the structure but does not go outside this range. ML can be relied upon to interpolate within this range. It has not been validated to extrapolate outside of this range. The review on AFW-HTC-CD papers discussed here suggests that it is still an open question whether ML can extrapolate.

ML models can be used in some cases for extensive experimentation or complex computational fluid dynamics simulations to predict the properties of nanofluids. One artificial neural network (ANN) model was able to predict the thermal conductivity and dynamic viscosity of nanofluids with an accuracy of R² > 0.95 by training with only the concentration of CD, the type of base fluid, and the working temperature^[9]. This result is specific to this dataset range but not a replacement for computational fluid dynamics or experimentation on any other dataset not tested. A small dataset for the laboratory testing typical of a new bio-refinery is best served by tree-based ensemble learning models such as random forest (RF) and extreme gradient boosting (XGBoost). The class of algorithms is better protected from overfitting^[29]. The methods give information on the most important feature (such as hydrothermal reaction parameter control of the surface heteroatom functionalization group). This has a much greater impact on the colloidal stability of the resulting CDs than the variations inherent in the raw materials^[29]. In addition to single-step predictions of properties, the evolution of computer-assisted design engineering points toward a new model: multi-objective optimization.

Table 3 gives an overview of recent literature and a cross-study map for a broad range of ML architectures applied to nanofluids, carbon precursors, target thermophysical or structural properties, and reported accuracies for those targets. The purpose is to show the engineering utility of these digital methods, not just their technical capabilities. In the last row, subcritical HTC of lignocellulosic biomass to hydrochar is included not as a carbon-dot characterization study, but as a process-level precedent for ML-guided HTC reactor optimization encompassing feedstock composition, reaction temperature, and residence time; accordingly, it does not evaluate carbon-dot synthesis or properties.

The accuracies in Table 3 are not directly comparable: reported n spans two orders of magnitude (20 to 2,500), and validation rigor varies from LOOCV on small samples to independent blind test sets on large ones. High R² at n < 100 (e.g., the active-learning study, n = 20-63) reflects efficient interpolation within a narrow, deliberately sampled region, not generalizable predictive power; only the large-sample rows (n > 1,000, external test sets) support a claim of robust performance. Reported R² values should therefore be read as an upper bound on in-domain accuracy, not as evidence of cross-precursor or cross-fluid transferability. Accordingly, Section 4 explores how these optimized CDs result in heat-transfer improvement of base fluids. Reported accuracies imply what was actually shown within each study’s training domain. It is assumed that future studies may eventually develop a precursor-to-property ML framework that can be used with any desired AFW feedstock, and it is outside the scope of this review. Finally, Figure 3 wraps up the flowchart by illustrating the full pathway from AFW feedstock through DoE/RSM experimental design generation to ANN, tree-ensemble, and hybrid ML models and ending at the final optimized nanofluid formulation.

The mechanisms above are listed in decreasing order of evidence. To date, none have been identified experimentally for AFW-derived CD nanofluids. Thus, they are stated as contributors, but without specifying relative importance. The enhancement of thermal conductivity (k) in CD-based nanofluids extends beyond classic effective medium descriptions to a combination of multi-scale thermodynamic processes.

Owing to their sub-10 nm, 0D structure, CDs exhibit rapid Brownian motion within the base fluid. It has been indicated that Brownian-driven micro-convection speeds up local heat and mass transfer^[36]. This mechanism is commonly used to explain the temperature dependence of the measured thermal conductivity. Rather than being a leading explanation, the estimates indicate that this contribution may be secondary; in addition, the speed of diffusion for the nanoparticles directly through the liquid is an order of magnitude slower than the rate of thermal diffusion. Furthermore, the same temperature dependence can also arise from effects related to aggregation-driven percolation and the temperature dependence of the interfacial liquid layer. Distinguishing between these requires experiments that separate the particle size from the volume fraction for matching surface chemistry, which have not been performed to date for CDs.

These oxygen-rich functional groups (-OH, -COOH) of biomass-derived CDs hydrogen-bond with polar base fluids to form a nanometer-thick ordered liquid layer at the solid-liquid interface that partially compensates for the phonon spectral mismatch and reduces the interfacial (Kapitza) resistance^[27]. This route relies on the correlations between surface oxygen content and enhancement and the results of molecular dynamics simulations of chemically similar functionalized carbon-water interfaces. However, it remains an unproven hypothesis in CD nanofluids. It would be desirable to confirm the mechanism directly via measurements of interfacial thermal conductance of AFW-derived CDs, e.g., using time-domain thermoreflectance. Yet, such measurements have yet to be performed. This mechanism can be considered to be only inferred^[27,38].

By adding 2D nanomaterials such as MXene into the formulation, CDs may act as intercalating nano-spacers that prevent restacking of the nanosheets and maintain a percolating thermal network^[35]. These results suggest an improvement in stability and interfacial interaction detected by Fourier Transform Infrared Spectroscopy (FTIR); however, it is important to note that no direct measurements of interlayer spacing were made under working conditions. The study also states that the enhancement in the case of the MXene/C-dot hybrid (42.2%) is less than that of pure MXene (50.0%)^[35], suggesting that in this system the process of hybridization sacrifices peak conductivity for dispersion stability. The true value of the 0D/2D architecture is the balance between conductivity and stability during its lifetime. This is the exact information that is lost during short-term bench testing.

Their unique optical properties such as their distinctive structural defects and surface states lead to a broad and strong absorption of light. These outstanding characteristics in photothermal conversion make them ideal candidates for direct absorption solar collectors, which can convert solar irradiation directly into local heat fast through non-radiative relaxation pathways^[27].

Table 4 lists the values of the reported thermal conductivity enhancement of CD- and carbon nanosphere (CNS)-based nanofluids to date according to the precursors used, synthesis methods, and operation conditions. Biomass-based CD/CNS nanofluids present the best thermal conductivity enhancement (up to 175%) with a hybrid 0D/2D structure indicating the best balance between high conductivity and stability.

One major challenge facing the commercialization of industrial nanofluids is long‐term suspension stability, particularly in severe high-temperature cycling conditions. In this respect, food waste‐derived CDs have been shown to greatly reduce this limitation due to their naturally self‐passivated surfaces. The reported stability window is short compared to the industrial life span of multiple years (up to ~ 37 days) ^[34], and data on accelerated aging under realistic thermal cycling is still lacking. Yet, the high absolute zeta potential (|ζ| > 30 mV, the conventional threshold for physical stability; |ζ| > 45 mV indicates excellent stability) provided by the high density of carboxyl and amino groups originating from the protein‐ and carbohydrate‐rich raw materials enables a strong electrostatic repulsion that leads to a high resistance to agglomeration even after prolonged heating-cooling cycles. Colloidal stability over 30 days has been reported for biomass‐derived CD nanofluids (e.g., 37 days without sedimentation)^[36]. This feature makes them an attractive alternative to conventional metallic nanoparticles for reducing clogging issues in heat-transfer loops.

In addition to these inherent advantages, CDs can be made suitable for operation in harsh environments through surface functionalization. Grafting with polyethylene glycol or ionic liquids can provide additional steric hindrance, locking colloidal stability. This modifies the rheology of the nanofluid toward non-Newtonian (shear-thinning) behavior, which should result in reduced pumping power compared to a Newtonian fluid with the same viscosity^[27,31]. However, neither source has calculated the resulting change in pumping power or pressure drop due to these changes. Thus, it is assumed this is an additional advantage based on their description of the rheology. Finally, a common finding in the reviewed papers is that there is mixed reporting of thermal conductivity enhancements and viscosity penalties; while many give a value for one parameter, few give a value for both.

None of the papers included in this review gives a comparison of pumping power or pressure drop for the nanofluid alongside this information. Therefore, it is currently impossible to calculate the overall thermohydraulic benefit of using AFW-derived CD nanofluids, the combination of the increase in heat-transfer rate and the associated increase in pumping power. This would control whether such systems can be used industrially. The published literature identifies this as a priority gap for future experimental work (Section CONCLUSION).

It is important to quantitatively show the “green” designation ascribed to biomass-derived CDs. Life cycle assessment (LCA), which complies with ISO 14040/14044, is a tool used to evaluate the impact on the environment from cradle-to-gate or cradle-to-grave. Fernandes et al.^[41,42] performed LCA by employing four different methodologies for life cycle impact assessment (LCIA) calculations, namely ReCiPe, Greenhouse Gas Protocol, USEtox, and AWARE. Their results show that using high‐yield hydrothermal methods does not provide any advantage in the environmental impact per gram when compared to low‐yield conventional protocols.

The main concept is that the chemical precursor (i.e., citric acid and urea) is the largest impact contributor to global warming, fine particulate matter formation, and ecotoxicity, followed by the electricity consumed in operating the reactor. This justifies the use of AFW as an alternative to refined chemical precursors in CD synthesis. Crista et al.^[43] replaced citric acid with spent coffee grounds in a one‐pot thermal carbonization reaction (200 °C, 4 h; 10 kWh per batch (batch mass not reported); product diameter 2.1-3.9 nm; quantum yield 2.9%-5.8%). Their results suggest that such replacement can decrease environmental impacts for several categories based on the ReCiPe 2016 midpoint. They ascribe it solely to the elimination of the refined chemical precursor. In a study involving nitrogen‐doped CDs synthesized hydrothermally from biomass‐based xylose and commercial xylose (180 °C, 4 h), Rodríguez-Carballo et al.^[44] concluded that the electric power and nitrogen precursor are both significant contributors to the total environmental impact. They did, however, highlight that the biomass‐based xylose decreases the global warming contribution per kilogram of the product. Across all three datasets, substituting refined chemical precursors with AFW streams is the single highest-leverage environmental intervention in CD synthesis, independent of reactor configuration.

If the system boundary is extended beyond the synthesis gate to include the whole life of the nanofluid, the benefit of AFW-derived CDs is further emphasized. In this context, Johny et al.^[45] prepared fluorescent CDs using Eucalyptus globulus leaves via one-pot HTC (quantum yield 60.7%; excitation/emission 320/445 nm) and applied a cradle-to-gate LCA approach. Although the HTC process did not constitute the greatest environmental impact, the presence of citric acid as a co-functionalization reagent caused the main impact. Accordingly, the lowest burden was obtained with a fully AFW-derived CD in the absence of any other synthetic co-reagents. Dolgun et al.^[46] measured the cumulative energy demand of the MWCNT/water and GNP/water nanofluids during their operational photovoltaic-thermal (PV/T) lifetime of 20 years.

It was shown that the improvements in thermal performance outweigh the greater energy inputs required for the synthesis of the nanofluids, and therefore, the resulting nanofluids are net-energy positive for their entire service life. This has also been highlighted by Sendão et al.^[47], who examined TiO₂-CD nanocomposites, whereby the substitution of the synthetic precursor for the CDs with a biomass-derived equivalent resulted in a change in the major environmental hotspot from the photocatalyst manufacture to the electricity generation component. These results are confirmed by recent reviews by Razzaq et al.^[48] on 30+ LCA studies carried out on nanofluids. Here, it was noted that up to 80% of the cradle-to-gate impact of nanofluids arises from the preparation of the NPs. However, once a long service phase is considered and waste-based materials replace pure refined feedstocks, the percentage contribution falls dramatically.

While this information is clear from the direction of the LCA results (Sections Life cycle assessment of agri-food waste-derived carbon dot synthesis and Environmental performance of CD-based nanofluids), there are still two factors that limit the quantitative reliability of the sustainability claim. First, none of the published LCA studies cover the whole lifecycle including the service life performance and/or the fate of the materials after their use phase has been included in an LCA analysis. To date, there is no published cradle-to-grave assessment of any AFW-derived CD nanofluid. Second, all the studies have used bench-scale batch autoclave data (< 1 L) as the sole source of process data (there is no LCI data available for larger scales). This is not representative of the actual industrial scale process, where continuous flow HTC is employed (> 100 L).

The change from bench-scale batch autoclave reactors to pilot-scale continuous flow HTC reactors influences the physical parameters of the HTC process. At bench scale, the autoclave reactions are carried out in near-isothermal conditions, whereas at pilot scale, due to the larger cross-sectional area, significant thermal gradients (15-30 °C) develop, leading to zones with lower temperatures (below the critical CD nucleation temperature) in the reactor^[6,7]. This will result in an increase in the size distribution, less overall graphitization, and lower fluorescence quantum yield than expected based on the small-batch autoclave reaction times. Malika et al.^[49] reported a similar trend for green synthesized iron-oxide nanoparticles, where they scaled up from 0.5 L to 20 L batches, which resulted in a 40% shift in the size distribution.

The lack of a reliable inventory compounds these issues. First, there is no standardized Life Cycle Inventory (LCI) data for HTC-specific unit operations, reactor pressurization, autogenous pressure maintenance, dialysis purification, or centrifugal separation. The lack of standardized LCI data for HTC unit operations is cited by De Matteis et al.^[50] as the major reason why comparative LCA for plant-derived CDs is difficult, with proxy data for chemically dissimilar processes potentially assigning the wrong dominant impact category altogether. There is also a strong publication bias in favor of reporting only synthesis runs where everything goes according to plan. Datasets from syntheses almost always omit failed runs and off-specification product. These will likely constitute the majority of data from an industrial quality-control record^[26]. A database compiled using this approach will underestimate the variability and the cost in terms of energy and materials associated with the process.

The most important knowledge gaps to be addressed include

• No cradle-to-grave LCA: Service-life thermal performance offsets (proven for MWCNT/water nanofluids over a 20-year PV/T lifespan by Dolgun et al.^[46]) have not yet been quantified for CD-based systems, meaning that their net contribution to environmental balance has not been determined.

• No pilot-scale HTC life cycle inventory: There is no available data for energy and mass balances of continuous-flow reactors. Batch-to-batch variability, heat-recovery efficiency, and cleaning-in-place water demand at scale are all unknown.

• Publication-biased training data: ML models and LCA inventories use only successful, optimized syntheses as training examples, leading to overestimates of expected yield and quality consistency because negative and sub-optimal results are not considered.

• Undefined end-of-life profile: No data on ecotoxicology, biodegradability, or recovery efficiency of dispersed CDs in spent heat-transfer fluid has been evaluated. This can impede industrial deployment in most jurisdictions.

Until such a data infrastructure is established, sustainability claims for AFW-derived CD thermal nanofluids remain directionally supported but insufficiently evidenced for regulatory or investment-grade decision-making. The four gaps identified above, publication-biased training data, absent pilot-scale HTC inventories, missing cradle-to-grave LCA, and an undefined end-of-life profile, are addressed by the consolidated research priorities in Section CONCLUSION. The research priorities identified here inform the recommendations in Section CONCLUSION directly.