2. Related work¶
Five features of the problem stated in §1 govern which methods may be applied to it: it is a mechanism-synthesis problem, so the literature on how such problems are posed governs the choice of objective; the objective is an application-level quantity, so the vehicle- and engine-modelling literature supplies the chain that prices geometry; the model is coupled, so the MDO-architecture literature governs how the disciplines are assembled; the gear pair is discrete, so the MINLP literature governs the outer loop; and the requirements are tolerance-limited, so the reliability literature governs how they are enforced. Each is reviewed below for what it settles and what it leaves open at this problem’s shape.
2.1 Extended-expansion engines and the mechanisms that realise them¶
That an expansion ratio larger than the compression ratio raises the ideal cycle efficiency has been known since Atkinson’s patents, and the modern literature is concerned with realising the asymmetry mechanically. Zhao [2017] separates the two families: valve-timed realisations (the Miller cycle, late or early inlet-valve closing), which shorten the effective compression stroke and lose displacement, and geometric realisations, in which a linkage gives the piston physically unequal strokes on alternate revolutions. The mechanism studied here belongs to the second family, whose canonical description is Watanabe et al. [2006] — the multiple-linkage general-purpose engine that became Honda’s EXlink, with a compression ratio of 8.5 against an expansion ratio of 12.3 and an indicated thermal efficiency raised from 27.3 % to 31.3 %. Dumböck et al. [2018] give an independent treatment of the same concept, and the physical model of §3.3 follows the standard treatment of engine friction, heat release and mean effective pressures in Heywood [1988].
Two things are settled by this literature and one is not. The thermodynamic benefit is real, and it is measured at a fixed mechanism: Watanabe et al. analyse a linkage whose dimensions are given rather than chosen. The half-speed shaft that produces the asymmetry does not change the firing frequency, power being taken from the crankshaft as usual. What is not settled is what the linkage costs: none of these studies carries the linkage dimensions through to the mass of the parts that must carry the inertia loads at speed, so none can say whether the thermodynamic gain survives the mechanism that delivers it. §5.4 carries it through.
2.2 Formulating a mechanism-synthesis problem¶
Mechanism synthesis divides into two families by what the optimizer is allowed to change. In the first the topology is given and only its dimensions are chosen; in the second the topology itself is an outcome of the optimization. The distinction matters here because the two families answer different questions about this engine, and because only one of them can be applied to it.
The first family — dimensional synthesis to a kinematic criterion — is the dominant formulation: choose link lengths to minimise a path or motion error, or to maximise a transmission quality index, subject to bounds on the envelope. Balli and Chand [2002] review the transmission angle as such an index and its use as both objective and constraint; Gosselin and Angeles [1990] give the Jacobian-based characterisation of the configurations where force transmission degenerates, which is what a transmission-angle criterion is a scalar proxy for. Cabrera et al. [2002] are representative of how these problems are solved: a genetic algorithm on a bounded box, the criterion evaluated kinematically.
The conventional treatment of this mechanism follows that pattern, maximising a lever-arm quality measure
subject to bounds on the two envelope dimensions \(H\) and \(B\), over the eleven linkage variables. Three limitations follow, and they belong to the formulation rather than to its solution.
It is multi-objective without an exchange rate. \(\eta\), \(H\) and \(B\) compete and nothing prices one against another, so the formulation yields a Pareto front and never a design. Weighted sums, \(\varepsilon\)-constraint and moving limits each produce a point, but the weights are the designer’s rather than the physics’.
Its central quantity is not an efficiency. With no friction in the model the virtual-work identity makes \(\int M_r\,\mathrm{d}\theta_1 \equiv \int P\,\mathrm{d}\lambda\) at every crank angle, so (2.1) is a ratio of two provably equal works: a kinematic quality measure in which nothing is lost. This is a general property of quasi-static transmission indices rather than a defect peculiar to this one — the literature treats them as proxies, and a proxy is only as good as its correlation with the quantity of interest.
It cannot see the parts. Nothing in a quasi-static formulation determines a cross-section, so the mechanism has no mass, no inertia loads and no speed dependence.
The alternative adopted here — carry the analysis through to the quantity the application scores — is standard practice in vehicle-level MDO rather than a novelty, and the ultra-efficiency-vehicle literature supplies the missing link: Gechev and Punov [2020] model an Eco-marathon prototype over its actual track and show that achievable consumption is set jointly by the powertrain map and the driving strategy, so neither can be scored without the other. That literature optimises the strategy at fixed hardware; what is optimised here is the hardware, with the strategy solved to optimality inside each evaluation.
Synthesising the topology, not only its dimensions¶
The second family generates the linkage itself. The line of work begun at Seoul National University by Kim et al. [2007] is the one relevant here: the design domain is filled with rigid blocks joined by zero-length springs of variable stiffness, and the stiffnesses are the design variables. A spring driven stiff welds two blocks into one rigid member; a spring driven soft becomes a revolute joint; so a single continuous parameterisation covers every linkage the domain can hold, and the topology, the joint positions and the link lengths are determined in one solve rather than assumed. The relaxation is the same device §2.4 describes for catalogue choices — a discrete present-or-absent decision made continuous and then penalised back to its extremes — applied to joints instead of to stock. The family has since been extended, in a form directly relevant to a mechanism whose 2:1 relation is carried by a gear pair, to planar gear-linkage mechanisms [Yim et al., 2019]; and a recent spring-connected link model [Tran and Jang, 2024] reaches the same answers with far fewer design variables.
Two things about it are worth stating precisely, because the springs invite a misreading. The result is a rigid-body linkage, not a compliant mechanism. The springs are a modelling device that the penalisation drives out; what is delivered is a set of rigid members and revolute joints. That is the reason the method belongs in this review at all. Compliant and soft mechanisms, which deliver their motion by elastic deformation, are not adapted to this application and are not considered anywhere in this study: the linkage carries kilonewton gas loads and joint reactions of 6 to 12 kN (§5.1), it must survive fatigue at a thousand cycles a minute, and an extended-expansion cycle depends on a kinematically exact 2:1 relation between the shafts — a relation that a deforming member does not hold, and whose error goes straight into the top-dead-centre gap that §5.2 shows this mechanism can least afford.
This study belongs to the first family: the EXlink topology is given, and eleven dimensions and a gear pair are chosen. That is a deliberate restriction rather than an oversight, and it has a cost the results make explicit — §5.4 measures what the topology is worth against a slider-crank and finds it cannot say whether a better topology exists, because two topologies establish a contrast and not a trend. The second family is the natural instrument for that question, and §5.6 applies it: a spring-connected synthesis run on this engine’s own requirements, with no mechanism assumed.
2.3 Architectures, derivatives, and thin feasible sets¶
Once mass, inertia and structural sizing enter, the model is coupled. Cramer et al. [1994] established the vocabulary — multidisciplinary feasible (MDF), individual discipline feasible (IDF), all-at-once — and Martins and Lambe [2013] give the modern survey, including the bi-level family of which Sobieszczanski-Sobieski et al. [2000] is canonical. MDF converges an MDA at every optimizer iteration, so every evaluated point is physically consistent and the design space stays small, at the cost of the inner iteration and of differentiating through it; IDF promotes the coupling variables to design variables with consistency constraints, at the cost of a design space that grows by the dimension of the coupling. The survey’s trade favours IDF when the MDA is expensive relative to the optimizer’s handling of a larger space; that trade is decided here by a single number, and §4.3 shows it is not close.
Whether a gradient method is admissible at all depends on the derivatives. Martins and Hwang [2013] unify the available routes and Sobieszczanski-Sobieski [1990] gives the global sensitivity equations that extend them to coupled systems. Finite differences are the default and are wrong here rather than merely inaccurate, several constraints being maxima over the crank revolution whose maximiser switches as the design moves. Complex-step [Martins et al., 2003] removes subtractive cancellation but not the switching problem; algorithmic differentiation [Griewank and Walther, 2008] is applicable at the cost of a dependency and of differentiating through the fixed-point solver unless taught not to. What is used here is analytic propagation combined with the envelope theorem — classical, Danskin [1966] for max-functions and Milgrom and Segal [2002] in general — which removes the switching problem rather than mitigating it (§4.2).
The geometry of the feasible set then selects the optimizer. Derivative-free direct search (Powell [1994], Nelder–Mead) and population methods (Deb et al. [2002]) proceed by sampling, and a sampling method evaluates points that lie off a measure-zero set with probability one, so it has no feasible point to improve from; §4.1 gives the measurement. Penalty and augmented-Lagrangian methods [Nocedal and Wright, 2006] are possible but inherit the conditioning. SQP with exact gradients [Kraft, 1988] moves along the manifold and is what is used. This is precisely why the synthesis literature above can use genetic algorithms and this study cannot: Cabrera et al. [2002] optimise over a box in which every sampled point is feasible. The same property restricts the escape from local optima, since multistart [Rinnooy Kan and Timmer, 1987], basin hopping and the global methods all assume the ability to generate a feasible starting point — leaving restarts constructed on the manifold (§4.6).
2.4 Discrete catalogue choices¶
The gear pair is discrete, so the problem is an MINLP; Belotti et al. [2013] survey the field. Exhaustive enumeration is exact on a small catalogue but offers no bound and no stopping criterion; relax-and-round is cheap but its rounded point generally leaves the feasible set; branch and bound needs a relaxation whose bound is meaningful. The two decomposition methods share a structure — a continuous NLP for fixed integers, a MILP master accumulating cuts — and differ in what the master is built from: optimal-value sensitivities for generalized Benders [Benders, 1962, Geoffrion, 1972], linearisations of \(f\) and \(g\) for outer approximation [Duran and Grossmann, 1986, Fletcher and Leyffer, 1994]. Fletcher and Leyffer [1994] also supply the treatment of infeasible sub-problems that a thin feasible set makes routine here. §4.4 chooses between them on which quantities this problem can supply reliably.
A separate strand treats catalogue choices by continuous relaxation with a penalty driving the relaxation back to a vertex — DMO [Stegmann and Lund, 2005] and SFP [Bruyneel, 2011], both descended from the SIMP penalty of Bendsøe [1989]. That machinery suits a large catalogue and a cheap sub-problem; here the catalogue has a handful of entries and each sub-problem solve is the expensive object, so it would convert an exactly solvable outer problem into an approximately solvable one.
2.5 Design under tolerance¶
The requirements of §3.1 are equalities that no manufactured part can meet exactly, so they must be relaxed to bands and the bands justified. Safety factors applied to the nominal constraint are simple, but the implied reliability is unknown and varies from constraint to constraint. Worst-case or fixed-margin robust design, \(g + k\sigma_g \le 0\), surveyed by Beyer and Sendhoff [2007], computes a margin rather than a probability, and is a reliability statement only under independence — which §4.5 shows fails here, correlations reaching \(\pm 1\) because the constraints are analytic functions of the same eleven dimensions. Reliability-based design optimization constrains a probability of failure: the first-order reliability method of Hasofer and Lind [1974], made algorithmic by Rackwitz and Fiessler [1978], gives \(\beta = -g/\sigma_g\) at the cost of one gradient, and its use inside an optimization is treated by Tu et al. [1999], by Du and Chen [2004], and in the survey of Valdebenito and Schüller [2010]. Because the failure event here is a system event, the union probability is needed; Ditlevsen [1979] gives the classical bounds and Genz [1992] the transformation that computes the correlated orthant directly. Tolerance classes are taken from [ISO, 2010].
2.6 Position of this work¶
This study is conventional in its machinery and unconventional only in what the machinery is applied to. The architecture comparison follows Martins and Lambe [2013], the analytic sensitivities Martins and Hwang [2013], the outer loop Duran and Grossmann [1986] and Fletcher and Leyffer [1994] as implemented in GEMSEO [Gallard et al., 2018], and the reliability treatment FORM [Hasofer and Lind, 1974, Rackwitz and Fiessler, 1978] with a system probability [Ditlevsen, 1979, Genz, 1992]. Nothing in §4 is a new algorithm.
What the literature leaves open is fivefold. The extended-expansion literature sizes no parts and so cannot price the mechanism against the cycle it enables. The dimensional-synthesis literature optimises a transmission quality that the virtual-work identity shows to be a proxy with no loss in it, and does so on a box where sampling methods are admissible. The topology-synthesis literature generates linkages against a kinematic target — a path, a motion, a timing — and stops there, so it too determines no cross-section and prices no part; a spring-connected block model asked for this engine would return a mechanism, not a range. The Eco-marathon literature optimises the strategy at fixed hardware. The RBDO literature treats the bounds as given data rather than asking which of them the mechanism is able to hold, and at what price in the objective. The contributions listed in §1.3 follow from joining these strands rather than from extending any one of them, and the problem is deliberately small enough that every claim is checked against a closed-form or sampled reference (§4.9).