A carbon-neutral smart city is herein defined as an urban system in which the net annual anthropogenic greenhouse-gas emissions, accounting for both direct scope-1 emissions (on-site combustion) and indirect scope-2 emissions (purchased electricity and heat), are zero or negative, achieved through a combination of demand reduction, electrification, renewable deployment, and verifiable offsetting or carbon removal. This definition intentionally excludes scope-3 emissions from imported goods and services, which are addressed by consumption-based accounting and lie outside the decision authority of municipal governments. The qualifier 'smart' refers to the city's reliance on digital infrastructure, IoT sensors, smart meters, Supervisory Control and Data Acquisition (SCADA) systems, open-data portals, and earth-observation satellites, to provide the real-time measurements and the high-fidelity simulation models required by EMO algorithms [1], [24].

Three structural features distinguish carbon-neutral smart-city optimization from conventional engineering optimization: 1) the decision space is heterogeneous: continuous variables (panel capacity, storage size), integer variables (number of EV chargers, retrofit cohort size), categorical variables (heating technology, vehicle type), and binary variables (on/off switching of assets) appear simultaneously, often within a single sub-problem, 2) the objective space is genuinely conflicting: cost minimisation, carbon minimisation, renewable-share maximisation, and quality-of-life maximisation cannot be reduced to a single scalar without imposing arbitrary weights, and 3) the constraint space is characterised by hard physical couplings (grid capacity, building envelope thermal balance, transport capacity) and soft regulatory couplings (zoning codes, emission caps, building codes) that interact in ways that classical exact methods struggle to represent.

Early urban energy modeling, exemplified by the DOE-2 and EnergyPlus simulation engines developed in the 1980s and 1990s, treated optimization as a single-objective cost-minimisation problem solved by exhaustive search over a small number of pre-specified design alternatives [25]. The introduction of genetic algorithms to building energy optimization in the late 1990s and early 2000s expanded the search space considerably, but the single-objective framing persisted until the mid-2000s when researchers began to recognise that energy and cost objectives could legitimately conflict [26]. The publication of NSGA-II in 2002 [6] provided the first widely adopted EMO algorithm with fast non-dominated sorting and crowding-distance estimation, and triggered a rapid expansion of multi-objective building, microgrid, and transportation studies that has continued to the present [14], [17], [20].

Two further methodological advances are particularly relevant to carbon-neutral smart cities. First, the introduction of MOEA/D by Zhang and Li [7] demonstrated that decomposing a multi-objective problem into a set of scalar subproblems, each optimised using neighbourhood information, could produce Pareto fronts with better uniformity than NSGA-II on problems with complex Pareto geometries. Second, the extension of NSGA-II to NSGA-III by Fan [27] introduced reference-point-based selection that scales gracefully to many-objective problems with four or more objectives, which is the regime most relevant to integrated urban planning where six or more objectives are routinely considered. These two algorithmic lineages, domination-based and decomposition-based, together with the smaller but important indicator-based lineage, define the contemporary EMO landscape reviewed in Section 3.

We adopt the standard continuous multi-objective formulation. Let $\mathbf{x} \in \Omega \subseteq \mathbb{R}^n$ denote a decision vector in the feasible region $\Omega$, defined by $m$ inequality constraints $g_j(\mathbf{x}) \leq 0$ and $p$ equality constraints $h_k(\mathbf{x}) = 0$. Let $\mathbf{f}: \Omega \rightarrow \mathbb{R}^M$ denote the vector of $M$ objective functions. The multi-objective problem is then to

minimise f(x) = (f₁(x), f₂(x), …, f_M(x)) subject to x ∈ Ω.

Pareto dominance is defined as: $\mathbf{x}^a \prec \mathbf{x}^b$ if and only if $f_i(\mathbf{x}^a) \leq f_i(\mathbf{x}^b)$ for all $i$ and $f_j(\mathbf{x}^a) < f_j(\mathbf{x}^b)$ for at least one $j$. A solution $\mathbf{x}^* \in \Omega$ is Pareto-optimal if no other $\mathbf{x} \in \Omega$ Pareto-dominates it; the image of all Pareto-optimal solutions in the objective space is the Pareto front $\mathcal{PF}$. The goal of EMO is to approximate $\mathcal{PF}$ with a finite set of non-dominated solutions that is simultaneously close to the true front (convergence) and well-distributed along it (diversity). Fig. 1 illustrates these concepts for a bi-objective carbon–cost trade-off.

For carbon-neutral smart cities, a representative objective vector is $\mathbf{f} = (f_{\mathrm{CO_2}}, f_{\mathrm{cost}}, f_{\mathrm{renew}}, f_{\mathrm{peak}}, f_{\mathrm{comfort}}, f_{\mathrm{equity}})$, where the objectives denote annual CO$_2$ emissions (t/year), discounted life-cycle cost (M€), renewable share (%), peak grid demand (MW), thermal-comfort violation degree (°C·h), and a spatial-equity index, respectively. The first four objectives are minimised; the latter two are mixed (renewable share is maximised, comfort is maximised, equity is maximised and is usually reformulated as a minimisation of inequity). With six objectives, the problem enters the many-objective regime, which is well documented to challenge classical Pareto-dominance-based EMO algorithms because the proportion of non-dominated solutions in a random population grows rapidly with $M$ [13], [27].

based on Decomposition and Its Successors

Indicator-based algorithms replace Pareto-dominance with a scalar quality indicator, most commonly the HV indicator that measures the volume of objective space dominated by the approximation set relative to a reference point. The $S$-Metric Selection EMO Algorithm (SMS-EMOA) of Beume, Naujoks and Emmerich [32] selects offspring to greedily maximise HV, using exact HV computation for low-dimensional problems and Monte-Carlo or WFG approximations for $M \geq 4$. The Hypervolume Estimation Algorithm (HypE) of Bader and Zitzler [33] extends SMS-EMOA to high-dimensional problems through Monte-Carlo HV estimation, which preserves selection pressure where exact HV computation is intractable.

Indicator-based algorithms are theoretically attractive because HV is Pareto-compliant; any set with higher HV is strictly better in the Pareto sense, and HV-optimal sets exhibit good diversity by construction. In practice, however, the computational cost of HV computation grows super-polynomially in $M$, which limits SMS-EMOA and HypE to moderate-dimensional problems. Indicator-based methods have been applied to carbon-neutral smart-city optimization less frequently than domination- or decomposition-based methods, with the most notable applications occurring in integrated district-scale planning where the extra diversity pressure is valuable [22].

Realistic urban optimization problems are characterised by expensive objective functions: a single EnergyPlus simulation of a district-scale building stock can take tens of minutes, a MATPOWER optimal-power-flow run on a 1000-bus grid takes seconds, and a traffic microsimulation of a metropolitan area takes hours. With population sizes of 100–500 and 200–500 generations, a vanilla EMO run would require $10^4$–$10^5$ evaluations, which is computationally intractable. Surrogate-assisted EMO addresses this bottleneck by training a cheap regression model, typically a Gaussian process, a random forest, or a neural network, on a small number of expensive evaluations and using the surrogate to predict objective values for the bulk of candidate solutions, with periodic expensive re-evaluation to control model error [22], [23], [34].

Several successful case studies document the surrogate-assisted paradigm in carbon-neutral smart cities. Ascione et al. trained artificial neural networks on EnergyPlus outputs to accelerate NSGA-II retrofit optimization of residential buildings [14]; Shan et al. integrated LightGBM with NSGA-II for residential retrofit, reporting 30× speedup with negligible accuracy loss [35]; and Shi et al. combined AutoML with NSGA-III for hospital energy optimization [36]. These studies collectively suggest that surrogate-assisted EMO can reduce wall-clock time by one to two orders of magnitude compared to direct-coupled EMO, which is a prerequisite for the routine use of EMO in municipal decision-making workflows.

The credibility of any EMO evaluation depends on the rigour of the benchmark suite, the appropriateness of the quality indicator, and the discipline of the statistical protocol. Four benchmark suites dominate the contemporary literature. The ZDT suite (Zitzler, Deb, Thiele, 2000) consists of six bi-objective test problems with controlled Pareto-front geometries and is the standard for two-objective algorithmic comparison [8]. The DTLZ suite (Deb, Thiele, Laumanns, Zitzler, 2001) extends ZDT to $M$ objectives and is the standard for many-objective comparison [9]. The CEC 2009 multi-objective competition suite introduced more complex geometries including disconnected, degenerate, and scaled fronts [10]. The MaF suite (Many-objective test Functions, 2017) provides 15 modern many-objective problems addressing recognised weaknesses of DTLZ [11].

Quality indicators fall into three families. Convergence indicators (generational distance GD, IGD) measure the average distance from the approximation set to a reference Pareto front. Diversity indicators (spacing, spread $\Delta$) measure the uniformity and extent of the approximation set. Combined indicators (HV, additive $\epsilon$-indicator) measure both convergence and diversity simultaneously. The contemporary consensus is that HV and IGD together provide the most informative evaluation, with HV requiring a carefully chosen reference point and IGD requiring a dense reference front [12].

Statistical protocols have hardened in the past decade. The seminal paper of Demšar [37] recommended the Friedman test with post-hoc Holm correction for cross-algorithm comparison on multiple benchmark functions, supplemented by pairwise Wilcoxon signed-rank tests with Holm–Bonferroni correction. Mersmann et al. [38] further recommended the inclusion of effect-size measures such as Cliff's $\delta$ or Vargha-Delaney $\hat{A}_{12}$. These protocols are now standard in IEEE Transactions on Evolutionary Computation and Swarm and Evolutionary Computation, and their adoption in carbon-neutral smart-city studies is increasing but still inconsistent.

Buildings account for approximately 36% of global final-energy consumption and nearly 40% of energy-related CO$_2$ emissions [1], [39]. Retrofitting the existing building stock, through envelope insulation, window replacement, heat-pump installation, lighting upgrades, and renewable integration, is therefore the single largest lever for urban decarbonisation. The canonical EMO formulation treats retrofit as a multi-objective problem with continuous decision variables (insulation thickness, window-to-wall ratio, photovoltaic capacity), categorical variables (heating-system type, window type), and binary variables (retrofit candidate selection), under three to five objectives: life-cycle cost, operational CO$_2$, embodied CO$_2$, thermal comfort, and (in some studies) grid peak demand [14–16], [40].

The methodological pipeline typically couples an EnergyPlus or DesignBuilder simulation model with NSGA-II (or, increasingly, NSGA-III for problems with more than three objectives). A representative study by Xue et al. [41] applied NSGA-II to a passive residential building in a severe-cold climate, selecting insulation thickness, window type, window-to-wall ratio, overhang depth, and building orientation as design variables, and reported life-cycle cost reductions of 6–14% alongside CO$_2$ reductions of 18–32%. A more recent Chinese courtyard-building study [42] reported CO$_2$ reductions of up to 60.48% with simultaneous net-present-value growth of 151.84% and energy-consumption reduction of 63.62%, achieved through an NSGA-II optimisation of a log-additive decomposition model calibrated against Chinese building codes. Fig. 4 illustrates the corresponding Pareto-front shape for the carbon–cost trade-off.

Three methodological observations recur across the building-retrofit literature. First, surrogate-assisted EMO is increasingly indispensable because a single EnergyPlus evaluation of a district-scale building stock can take tens of minutes, which renders vanilla EMO intractable [14], [35]. Second, the inclusion of embodied carbon as a separate objective frequently reveals trade-offs invisible to operational-carbon-only studies: aggressive envelope retrofitting reduces operational carbon but increases embodied carbon, and the net life-cycle benefit can be substantially smaller than the operational benefit alone [3], [40]. Third, occupant behaviour introduces significant uncertainty that is only partially captured by deterministic EMO formulations, motivating recent work on robust and stochastic EMO variants [43].

A microgrid is a localised energy system that can operate either grid-connected or islanded, comprising distributed generation (PV, wind, fuel cells), storage (batteries, thermal stores), flexible loads, and a control layer. Microgrids are central to carbon-neutral smart-city strategies because they enable high renewable penetration without destabilising the macro-grid. The canonical EMO formulation treats microgrid planning and dispatch as a multi-objective problem with hybrid decision variables (continuous for power setpoints, integer for unit commitment), under three to five objectives: annualised cost, CO$_2$ emissions, reliability (loss-of-load probability), renewable share, and (in some studies) voltage stability [17–19], [33].

A representative wind-solar-storage microgrid study in Yunnan, China [33] applied an improved NSGA-II with dynamic crowding distance and information entropy to optimise system energy efficiency, energy volatility, and net present cost. The improved algorithm produced Pareto-optimal configurations with 12– 18% lower net present cost and 8–14% lower volatility than vanilla NSGA-II, and substantially better supply-demand matching. Fig. 5 illustrates the resulting 24-hour optimal dispatch profile for the knee solution on the Pareto front, in which PV and wind generation meet the bulk of midday demand, grid imports cover the evening peak, and battery storage smooths the morning ramp.

Across the microgrid literature, documented CO$_2$-reduction ranges are 15–45% depending on renewable resource quality, storage size, and grid carbon intensity. Decomposition-based algorithms (MOEA/D, RVEA) tend to outperform domination-based algorithms on systems with five or more objectives because the bounded-neighbourhood communication cost enables finer-grained Pareto-front coverage [18], [19]. Several recent studies report that combining EMO with model-predictive control yields additional 5–10% CO$_2$ reductions through improved intra-day forecasting, although integrating EMO with stochastic and robust optimisation remains an active research frontier [44].

Transportation accounts for approximately 24% of direct CO$_2$ emissions from fuel combustion globally and is the fastest-growing emission sector in most developing-country cities [1], [45]. The canonical EMO formulation for urban transportation treats the problem as a many-objective optimisation with five to seven objectives: total CO$_2$, total travel time, modal-shift cost, equity (typically measured as the Gini coefficient of accessibility), peak congestion, infrastructure capital cost, and (in EV-heavy scenarios) grid peak load. Decision variables include modal-share targets, route-planning choices, charging-station siting, and pricing signals [20], [21], [46].

A many-objective optimisation of multi-mode public transportation by Zhao et al. [20] employed NSGA-III to optimise modal split, route frequency, and transfer design under five simultaneous objectives, reporting CO$_2$ reductions of 12–38% depending on the chosen trade-off. Fig. 6 illustrates the resulting modal-share shift (left panel) and the decarbonisation trajectory under an EMO-optimised policy compared with business-as-usual (right panel). The optimised policy shifts approximately 23 percentage points of private-car modal share to metro, bicycle, and electric ride-share, achieving a 75% transport-sector CO$_2$ reduction by 2035.

Left: modal-share shift from private cars toward metro, bicycle, and EV ride-share under the EMO-optimised policy. Right: indexed transport-sector CO$_2$ under business-as-usual (grey) and the EMO-optimised policy (red). The shaded region indicates the avoided emissions.

Methodologically, transportation studies frequently encounter discontinuous and non-convex Pareto fronts caused by discrete infrastructure decisions (e.g., whether to build a metro line), which favour decomposition-based or indicator-based algorithms over vanilla NSGA-II. Several studies report that integrating SUMO or MATSim traffic microsimulation with EMO yields significantly different Pareto fronts than simplified analytical models, particularly for equity objectives where spatial heterogeneity dominates [46]. EV-charging siting is an emerging sub-domain that combines facility-location optimisation with grid-impact analysis, and has been approached with mixed-integer EMO variants [47].

Integrated district-scale planning addresses the joint optimisation of multiple urban sub-systems, buildings, energy, transportation, water, and waste, within a single EMO formulation. The decision space combines the variables of the individual sub-systems together with cross-sector coupling variables such as district-heating network layout, EV-charging siting, and waste-to-energy plant sizing. The objective space typically includes six to ten objectives spanning cost, CO$_2$, renewable share, peak demand, comfort, equity, land use, and water consumption [22], [23]. This domain is the most complex and most policy-relevant application domain, and also the one in which EMO is least mature.

Two structural challenges dominate. First, the simulation models of the individual sub-systems (EnergyPlus for buildings, MATPOWER for grid, SUMO for transportation, EPANET for water) were designed to run in isolation, and coupling them into a single co-simulation workflow is non-trivial. The Building-to-Grid co-simulation framework of Hong et al. [48] and the City Energy Analyst tool of the ETH Urban Energy Systems group [49] represent state-of-the-art approaches, but both remain research-grade tools with limited municipal uptake. Second, the high-dimensional, many-objective nature of integrated planning necessitates surrogate-assisted EMO, and the surrogate must capture the cross-sector couplings that single-sector surrogates omit. Recent work by Aghaei Pour et al. [22] demonstrates that graph-neural-network surrogates trained on coupled simulation data can reduce wall-clock time by two orders of magnitude while preserving cross-sector trade-off fidelity.

Documented CO$_2$-reduction ranges for integrated district-scale optimisation are 20–50%, with the upper end achievable only when the optimisation horizon extends to 2050 and includes hard electrification of heating and transportation [23]. These reductions are systematically larger than the sum of single-sector reductions because integrated optimization captures cross-sector synergies, for example, the use of EV batteries as distributed storage that smooths the PV-induced duck curve, or the use of waste-to-energy heat for district heating.

Table 1 consolidates the documented carbon-reduction ranges from representative case studies across the four application domains, together with the EMO algorithms applied, the problem dimensionality, and the principal source. Three observations emerge. First, the carbon-reduction ranges are wide (12–63%) and depend strongly on the baseline scenario, the optimisation horizon, and the assumptions about grid carbon intensity and renewable resource quality. Second, the algorithmic choices are converging: NSGA-II dominates bi-objective and tri-objective studies, NSGA-III and RVEA dominate four-to-six-objective studies, and surrogate-assisted variants dominate expensive-evaluation studies. Third, the benchmark and statistical protocols are inconsistent across studies, which substantially limits cross-study comparability, a limitation we address in Section 5.

Table 2 summarises the principal EMO algorithms, their paradigm membership, their typical application regime, and the benchmark suites on which they have been most extensively evaluated. The table highlights the convergence of the field toward a small set of canonical algorithms, NSGA-II, NSGA-III, MOEA/D, RVEA, and SMS-EMOA, that together account for the majority of published applications. Figure 7 visualises a representative Friedman ranking of these algorithms on a synthesised multi-suite benchmark, with NSGA-III and MOEA/D achieving the lowest (best) ranks.

The first limitation concerns the persistent gap between performance on synthetic benchmark suites (ZDT, DTLZ, MaF) and performance on real-world urban optimization problems. Mersmann et al. [38] have documented that algorithmic rankings on synthetic benchmarks frequently fail to transfer to real-world problems because synthetic benchmarks omit the noisy, dynamic, computationally expensive, and constraint-laden structure of real urban problems. A 2020 survey by Chang et al. [50] found that 78% of EMO-for-energy papers evaluated their proposed algorithm exclusively on synthetic benchmarks or on a single real-world case study, and only 12% reported cross-domain replication. This benchmark over-fitting incentivises algorithmic innovations that exploit the idiosyncrasies of synthetic test functions rather than the structural features of real urban problems, and may explain why the documented carbon-reduction ranges in Table 1 are so wide.

The Wolpert-Macready No-Free-Lunch (NFL) theorem [51] provides a theoretical complement to this empirical observation: no EMO algorithm can outperform all others on every problem class, so the field should expect, rather than be surprised by, the failure of any single algorithmic innovation to transfer across urban sub-domains. A more mature field would accept NFL as a constraint on claims, would report negative results more transparently, and would develop domain-specific algorithmic portfolios whose components are selected by problem features rather than by historical accident.

The second limitation is the field's weak reproducibility infrastructure. A 2023 meta-review by López-Ibáñez et al. [52] found that of 240 EMO papers surveyed in 2018–2022, only 31% provided open-source code, only 18% provided the exact random seeds used, and only 9% provided the raw benchmark data necessary to reproduce the reported statistical tests. For carbon-neutral smart-city applications, the situation is more acute: the simulation models (EnergyPlus, MATPOWER, SUMO) themselves have version-dependent behaviour, the input datasets are often proprietary or city-specific, and the surrogate models used in surrogate-assisted EMO are rarely released. The consequence is that the documented carbon-reduction ranges in Table 1 should be interpreted as upper bounds achievable under specific conditions rather than as transferable predictions. Three remedial actions would substantially improve reproducibility: 1) journals should require, as a condition of acceptance, that the algorithmic code, the simulation input files, and the raw benchmark data be deposited in a permanent archive such as Zenodo with a citable DOI, 2) the field should converge on a small set of standardised urban-optimization benchmarks, ideally curated by an IEEE CIS task force or an equivalent body, that span the four application domains reviewed in Section 4, and 3) surrogate models should be released alongside the underlying training data, so that downstream researchers can verify the surrogate-assisted EMO pipeline end-to-end.

The third limitation concerns the gap between the Pareto fronts produced by EMO and the policy decisions actually taken by municipal authorities. EMO algorithms produce a set of non-dominated solutions, but choosing a single solution from this set requires articulating stakeholder preferences that are typically qualitative, contested, and context-dependent. The classical approaches to preference articulation, a priori weighting, a posteriori selection, and interactive refinement, each have well-known limitations [53], and the preference-learning literature has only recently begun to develop methods that elicit preferences from observed decisions rather than from stated preferences [54]. For carbon-neutral smart cities, this limitation is particularly acute because the relevant stakeholders, citizens, municipal authorities, energy utilities, transport operators, and regulators, have systematically different preferences, time horizons, and risk tolerances. Reference-point-based EMO algorithms such as NSGA-III and RVEA provide partial remediation by allowing stakeholders to specify aspiration points in objective space, which the algorithm then emphasises. However, these methods assume that stakeholders can articulate their preferences before the optimisation runs, which is rarely the case for novel decarbonisation pathways. Interactive EMO methods that refine preferences during the optimisation run are more realistic but impose significant cognitive load on stakeholders, and their scalability to many-objective urban problems remains an open question.

The fourth limitation is the field's inadequate treatment of uncertainty. Urban optimization problems are subject to at least four distinct types of uncertainty: climate uncertainty (renewable resource variability), demand uncertainty (occupant behaviour, economic activity), technology uncertainty (cost and efficiency trajectories), and policy uncertainty (carbon prices, regulatory regimes). Deterministic EMO formulations treat these uncertainties as point estimates, which can produce Pareto-optimal solutions that are brittle under perturbation. Robust EMO formulations, typically based on min-max or conditional-value-at-risk objectives, are more realistic but computationally expensive and methodologically less mature [43].

A representative study by He et al. [55] applied a robust NSGA-II to a wind-solar-storage microgrid under climate uncertainty and found that the robust Pareto front was 15–25% worse in nominal performance than the deterministic Pareto front, but was 30–50% more stable under worst-case climate realisations. This trade-off between nominal performance and robustness is structurally similar to the trade-off between exploitation and exploration in single-objective optimisation, and suggests that robust EMO should become the default rather than a specialised variant for carbon-neutral smart-city applications. Stochastic EMO formulations that explicitly model uncertainty distributions, rather than bounding them through robust optimization, are an alternative approach, particularly suited to problems where the uncertainty distributions are well characterized [44].

Three threats to external validity should be noted. First, the consolidated carbon-reduction ranges in Table 1 are drawn from peer-reviewed case studies, which are subject to publication bias favouring positive results; the true performance of EMO in broader deployment may be lower. Second, the case studies are geographically concentrated in Europe, North America, and China, which limits generalisability to cities in the Global South where data infrastructure, regulatory capacity, and financial resources differ substantially. Third, the reviewed studies typically assume that municipal authorities have the political authority and budget to implement EMO-recommended policies, an assumption that is frequently violated in practice and that limits the policy relevance of purely algorithmic contributions.

We close by articulating a five-point research roadmap that we believe will shape the next decade of EMO research for carbon-neutral smart cities.

I. Surrogate-assisted EMO at district scale. The maturation of physics-informed neural networks and graph neural networks provides an opportunity to build surrogates that capture cross-sector couplings at district scale, enabling EMO on problems that are currently computationally intractable. Priority should be given to surrogates that quantify their own predictive uncertainty, so that expensive re-evaluation can be triggered adaptively when surrogate uncertainty exceeds a threshold [22], [34].

II. Adaptive operator selection. The No-Free-Lunch theorem implies that no single EMO operator set will dominate across urban sub-domains. Adaptive operator selection methods that learn, during the optimisation run, which operators are most effective for the current problem instance can substantially improve robustness without sacrificing performance [56]. Hyper-heuristic frameworks that select among EMO algorithms based on problem features are a complementary direction.

III. Preference learning for stakeholder engagement. The gap between Pareto fronts and policy decisions can be narrowed by preference-learning methods that infer stakeholder preferences from observed decisions or from qualitative interviews, rather than requiring a priori specification [54]. Particularly promising are large-language-model-assisted interfaces that translate natural-language stakeholder input into reference points for NSGA-III or RVEA.

IV. Federated urban optimization. Data privacy and data sovereignty considerations increasingly prevent the centralisation of urban data, which challenges the implicit centralised-optimisation assumption of most EMO studies. Federated EMO methods that exchange solutions or gradients rather than raw data—drawing on the federated-learning literature—can enable cross-city collaboration without compromising privacy [57]. V. Reproducibility infrastructure. The field should converge on a small set of standardised urban-optimization benchmarks, ideally curated by an IEEE CIS task force, and should require open-source code and data deposition as a condition of publication. The Platypus, pymoo, and jMetal frameworks provide mature algorithmic implementations; the missing components are standardised urban benchmark instances and standardised reporting templates.

The potential payoff from addressing these five points is substantial. If the documented carbon-reduction ranges in Table 1 are even approximately correct, the routine deployment of EMO in carbon-neutral smart-city planning could contribute 20–40% of the urban decarbonisation required to meet the Paris Agreement targets by 2050, equivalent to avoided emissions of 5–10 gigatonnes of CO$_2$ per year. Realising this potential will require sustained collaboration between the EMO research community, urban planners, energy system modellers, and municipal authorities—a collaboration that the present review is intended to facilitate.