Semi-analytical Solver (SAS) Overview

• Semi-analytical solvers are methods that blend closed-form analytic representations with localized numerical evaluations to solve differential-algebraic and PDE systems across various domains.
• They derive truncated series or polynomial expansions offline and use recursive or averaging techniques online to reduce computation and enhance parallelism.
• Applications span power system transient stability, orbit propagation, natural-gas networks, elasticity, and shock problems, yielding significant speedups and tailored error control.

Searching arXiv for the cited SAS-related papers and closely related formulations. Search query: arXiv (Duan et al., 2014) semi-analytic solutions power system differential algebraic equations; (Liu et al., 2018) differential transformation method power system; (Wang et al., 2018) time-power series SAS; (Amato et al., 2018) semi-analytical orbit propagation STELA; (Xu et al., 2022) natural gas network semi-analytical solution; (Ciro et al., 2014) Grad-Shafranov semi-analytical solver; (Zhu et al., 18 Jun 2025) semi-analytical forward solver elastic parameters; (Gerberding et al., 20 Sep 2025) Noh problem semi-analytic solutions. A semi-analytical solver (SAS) is a solver architecture in which part of a governing model is handled in closed or semi-closed form while the remaining computation is carried out by localized numerical evaluation, reduced auxiliary solves, or short-window propagation. In the power-system formulation of SAS, the semi-analytic solution is a closed-form function of symbolic variables including time, the initial state and the parameters on system operating conditions, and hence able to directly give trajectories on system state variables that are accurate for at least a certain time window (Duan et al., 2014). In orbital mechanics, semi-analytical methods remove fast oscillations analytically and integrate only the slow evolution of mean elements (Amato et al., 2018). Other arXiv uses of the term include multivariate-polynomial cellwise solutions for natural-gas PDEs (Xu et al., 2022), eigenfunction-based semi-analytical equilibrium construction for the Grad–Shafranov equation (Ciro et al., 2014), a direct method of lines that preserves analytical radial dependence in elasticity (Zhu et al., 18 Jun 2025), and self-similar shock solutions closed by Rankine–Hugoniot algebraic systems for the Noh problem with a black-box equation of state (Gerberding et al., 20 Sep 2025).

1. Definitional scope and common structure

The phrase “semi-analytical solver” does not denote a single algorithm. In the cited literature it designates a class of methods that retain an analytic representation for a strategically chosen part of the problem while delegating the rest to recursion, averaging, collocation, linear algebra, or localized numerical integration. In power systems this analytic part is typically a short-time series in time, parameterized by initial conditions and operating-point parameters (Duan et al., 2014). In orbit propagation it is the analytical removal and later recovery of short-periodic terms from osculating elements (Amato et al., 2018). In elasticity it is the analytical solution along one coordinate after discretization in another (Zhu et al., 18 Jun 2025).

Taken together, these formulations suggest a common structural pattern. First, the governing equations are rewritten so that one component of the dynamics becomes symbolically tractable: a decomposition series, a mean-element system, a polynomial ansatz, a separable eigenexpansion, or a self-similar reduction. Second, the semi-analytical representation is truncated or specialized to a finite basis. Third, the resulting object is evaluated or fitted over finite windows, cells, or regions. Fourth, auxiliary mechanisms control validity: divergence indicators in power-system transient simulation, resonance-retention criteria and perturbation truncation in orbit propagation, collocation and continuity conditions in gas networks, boundedness and eigenvalue selection in elasticity, or Newton solves for Rankine–Hugoniot closure in the Noh problem [(Duan et al., 2014); (Amato et al., 2018); (Xu et al., 2022); (Zhu et al., 18 Jun 2025); (Gerberding et al., 20 Sep 2025)].

A plausible implication is that “semi-analytical” in current computational practice is best understood operationally rather than taxonomically: it identifies a solver in which analytic structure is preserved only to the extent that it yields a computational advantage without requiring a fully closed-form global solution.

2. Power-system SAS formulations

In transient stability simulation, the baseline model is a nonlinear differential-algebraic system

$\dot x(t)=f(x(t),y(t),p), \qquad 0=g(x(t),y(t),p),$

with $x(t)\in\mathbb{R}^m$ the dynamic states, $y(t)\in\mathbb{R}^k$ the algebraic variables, and $p$ the fixed parameters such as inertia $H$, damping $D$, machine EMFs, network admittances, and mechanical power inputs (Duan et al., 2014). In practice one often eliminates $y$ via network equations to obtain coupled ODEs. For the classical 2nd-order model of a $K$-machine system,

$M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$

or, in first-order form,

$$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$

The Adomian-decomposition formulation derives the SAS by expanding both the state trajectory and the nonlinear term in infinite series and then obtaining each term recursively in closed form (Duan et al., 2014). For an ODE of the form $x(t)\in\mathbb{R}^m$0, the method applies the Laplace transform, assumes

$x(t)\in\mathbb{R}^m$1

with Adomian polynomials

$x(t)\in\mathbb{R}^m$2

and derives recursive relations in the Laplace domain. Truncation after $x(t)\in\mathbb{R}^m$3 terms yields

$x(t)\in\mathbb{R}^m$4

which is accurate over a finite time window $x(t)\in\mathbb{R}^m$5 that depends on $x(t)\in\mathbb{R}^m$6 and on system parameters (Duan et al., 2014). A modified ADM may fold linear damping terms and certain constant parts into the zeroth term, so that each $x(t)\in\mathbb{R}^m$7 becomes a polynomial in $x(t)\in\mathbb{R}^m$8 plus possibly sinusoids.

The differential transformation method (DTM) produces the same general short-time objective through a different coefficient calculus (Liu et al., 2018). For an analytic function $x(t)\in\mathbb{R}^m$9, the transform and inverse transform are

$y(t)\in\mathbb{R}^k$0

Applied to an index-0 DAE,

$y(t)\in\mathbb{R}^k$1

the DTM yields recursions

$y(t)\in\mathbb{R}^k$2

followed by the truncated expansions

$y(t)\in\mathbb{R}^k$3

The distinguishing claim is computational: the recursive formula avoids generating and storing its complete symbolic expression, which makes both derivation and evaluation of the SAS more efficient especially for multi-machine power systems (Liu et al., 2018).

A third power-system variant writes the SAS directly as an $y(t)\in\mathbb{R}^k$4th-order power series about $y(t)\in\mathbb{R}^k$5,

$y(t)\in\mathbb{R}^k$6

with coefficients obtained by matching the Taylor expansion of $y(t)\in\mathbb{R}^k$7 and the recursion

$y(t)\in\mathbb{R}^k$8

(Wang et al., 2018). This framework extends to general DAEs through the dynamic-bus method, in which every PV or load bus is represented as a dynamic bus with voltage series in time, component SASs are evaluated in parallel, the network algebraic equations are solved once at the end of the window, and the voltage-series coefficients are re-fitted over the interval (Wang et al., 2018).

3. Windowing, offline–online execution, and error control

The power-system literature gives the SAS a distinct offline–online decomposition. In the two-stage SAS-based approach, the offline stage selects symbolic variables such as time, initial states, parameters, and optionally selected admittance matrix entries to cover multiple contingency topologies; applies Multi-stage Modified ADM to derive symbolically $y(t)\in\mathbb{R}^k$9 up to order $p$0 for each generator; stores the closed-form expressions; and estimates the maximum length of window of accuracy $p$1 (Duan et al., 2014). The online stage carries out numerical integration only through the fault-on period to obtain the post-fault initial state, evaluates the SAS for all generators over the first interval, extracts the terminal state, and repeats sequentially until the horizon is covered.

Local validity is central. Because a truncated SAS departs from the true solution beyond some $p$2, the power-system ADM paper defines a divergence indicator

$p$3

and chooses a threshold $p$4 so that $p$5 (Duan et al., 2014). In a multi-machine case one computes $p$6 per generator and then picks $p$7. The time-power-series formulation uses an error-rate upper bound

$p$8

with cumulative error bounded by $p$9, and selects the largest $H$0 satisfying $H$1 (Wang et al., 2018). These are two different mechanisms for the same operational requirement: finite-window reliability.

The DTM formulation makes the offline–online split even more explicit. Offline, one derives once the recursive rules symbolically in terms of the initial state and parameters. At runtime, whenever a new contingency or operating point changes the initial state or parameters, the same recursion rapidly computes the numeric coefficients for the current window (Liu et al., 2018). The paper further studies an optimal order $H$2 by balancing increased window length $H$3 against increased per-window evaluation time, with total time

$H$4

The reported example on the IEEE 39-bus system states that, for tolerance $H$5 pu, $H$6 and $H$7 s were optimal (Liu et al., 2018).

A recurrent misconception is to equate SAS with a single long-horizon closed-form trajectory. The cited formulations do not support that interpretation. In these power-system papers, the semi-analytic object is explicitly local, and long simulations are obtained by repeated reinitialization over short windows [(Duan et al., 2014); (Liu et al., 2018); (Wang et al., 2018)].

4. SAS in orbit propagation, gas networks, plasma equilibrium, elasticity, and shock verification

In Earth-orbit propagation, semi-analytical methods seek to retain only the slow evolution of orbital elements by removing fast orbital-periodic oscillations analytically (Amato et al., 2018). The state is written in a convenient element set, for example the non-singular equinoctial elements

$H$8

with

$H$9

Osculating equations are written in Gauss form, perturbations are expanded in Fourier and Legendre series, short-periodic corrections are separated from mean elements,

$D$0

the mean-element rates are obtained by averaging over the fast angle or angles, and the averaged equations are then integrated with a numerical solver (Amato et al., 2018). The computational benefit is that, by eliminating the $D$1-revolution oscillations analytically, the numerical solver may take much larger steps, on the order of days.

In natural-gas transients, SAS is a cellwise polynomial approximation for the isothermal pipeline PDE system

$D$2

(Xu et al., 2022). The $D$3 domain is partitioned into uniform cells of size $D$4, coordinates are normalized to $D$5, and in each cell the SAS ansatz is

$D$6

$D$7

where the coefficients are determined by PDE residual equations, initial-boundary projection at collocation points, and seamless coupling between adjacent cells. Each time slice reduces to a square linear system $D$8, and a reduced scheme, SAS-2, replaces the friction term by a cell-wise linear model,

$D$9

to simplify assembly and reduce run time with minor loss of accuracy (Xu et al., 2022).

For the Grad–Shafranov equation, the solver becomes semi-analytical after specific constitutive assumptions render the PDE linear (Ciro et al., 2014). With pressure proportional to the poloidal flux and $y$0 quadratic in $y$1, the normalized equation becomes

$y$2

which is decomposed into a particular solution and a homogeneous part with eigenvalue parameter $y$3 (Ciro et al., 2014). The full flux is then written as a linear combination of independent basis solutions whose coefficients are fitted under physical constraints such as aspect ratio, elongation, triangularity, X-point geometry, and current-profile conditions.

In heterogeneous and anisotropic elasticity, the semi-analytical forward solver is the direct method of lines (Zhu et al., 18 Jun 2025). A star-shaped domain is mapped to log-polar coordinates,

$y$4

only the angular variable is discretized with a periodic $y$5 finite-element space, and the static Navier equations reduce to a second-order ODE system in $y$6,

$y$7

Seeking modes of the form

$y$8

yields a quadratic eigenvalue problem

$y$9

and the bounded solution is reconstructed from the eigenmodes with nonnegative real part (Zhu et al., 18 Jun 2025). The paper emphasizes that the eigen-exponents converge to the true stress-singularity exponents and that no special singular elements are required.

For the Noh problem with a black-box equation of state, the semi-analytic object is a self-similar two-region shock solution (Gerberding et al., 20 Sep 2025). The pre-shock solution is

$K$0

the post-shock state is constant with $K$1, and Rankine–Hugoniot conditions reduce the closure to a nonlinear $K$2 system in $K$3 or $K$4 depending on the EoS representation. A Newton iteration then produces the shock speed and post-shock state, from which the full piecewise semi-analytic solution follows (Gerberding et al., 20 Sep 2025).

5. Computational characteristics and reported performance

The principal computational claim of SAS methods is not that they eliminate numerics, but that they shift expensive computation away from the most frequent operations. In the ADM-based power-system solver, once offline derivation is done, online costs are only closed-form evaluations and no iterative solves; the method also exhibits a high degree of parallelism because each generator is independent (Duan et al., 2014). In the DTM formulation, the avoidance of complete symbolic expansion reduces derivation and storage burdens, and coefficients at the same order but different machines are independent, which enables parallel evaluation (Liu et al., 2018). In the time-power-series formulation, the computational advantage comes from evaluating component-level SASs over large windows while solving the network algebraic equations far fewer times (Wang et al., 2018).

Several papers report representative benchmarks:

Domain	Representative setup	Reported outcome
Power systems (ADM)	IEEE 10-machine, 39-bus; order $K$5; $K$6 s	online $K$7 ms per 4 s simulation in Case A; speedup $K$8 real time (Duan et al., 2014)
Power systems (DTM)	IEEE 10-machine 39-bus; $K$9; $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$0 s	serial SAS time $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$1 s; ideal 40-core time $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$2 s (Liu et al., 2018)
Power systems (time-power series)	Polish 2383-bus, 327 machines + 366 motors	ideal component-level parallelism gives total $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$3 s; $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$4 speed-up vs. Euler (Wang et al., 2018)
Orbit propagation	STELA vs THALASSA over decades	STELA $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$5 faster for LEO at $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$6 km accuracy; non-averaged methods comparable or faster in GTO/HEO (Amato et al., 2018)
Natural gas	6-node network, $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$7 s	SAS-2 with $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$8 gives $M\ddot\delta + D\dot\delta = P_m - P_e(\delta),$9 in $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$0 s (Xu et al., 2022)

The orbit-propagation comparison is especially instructive because it isolates the cost of the semi-analytical approximation itself. The paper decomposes total integration error into dynamical, model truncation, short-periodic, and numerical components, and argues that the first three are attributable to distinct approximations required by the method of averaging (Amato et al., 2018). As a result, STELA can be very fast in LEO and MEO regimes, but its attainable accuracy is fundamentally limited by averaging-inherent errors. For the test cases considered, regularized non-averaged methods were found to be up to two times slower than semi-analytical for the LEO orbit, to have comparable speed for the GTO, and to be ten times as fast for the HEO for the same accuracy (Amato et al., 2018).

The natural-gas results show a different mechanism. There the gain comes from using a much larger grid cell than the widely used finite difference method. In the single-pipeline example, SAS-2 with $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$1 reaches error $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$2 in $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$3 s, whereas the reference FDM with $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$4 m and $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$5 s takes $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$6 s to simulate $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$7 s (Xu et al., 2022). This is not merely a faster linear solve; it is a different discretization regime enabled by the cellwise polynomial ansatz.

6. Limitations, misconceptions, and methodological significance

The literature is explicit that SAS methods have domain-specific limits. In power systems, the SAS is only locally accurate; one must choose the window carefully or adaptively monitor the divergence indicator, and high-order SAS increases expression complexity (Duan et al., 2014). Offline symbolic derivation can also become expensive as the number of symbolic parameters grows. The time-power-series formulation similarly notes that offline derivation can be memory- and CPU-intensive for very large-order series and that network algebraic equations still have to be solved, even if much fewer times (Wang et al., 2018). In the DTM setting, larger order increases both the admissible time window and the evaluation burden, so optimal performance requires order selection rather than monotonically increasing truncation order (Liu et al., 2018).

In orbit propagation, the main limitation is structural rather than implementation-specific. Dynamical error, model truncation error, and short-periodic reconstruction error cannot be reduced by simply decreasing the numerical step size; one must derive higher-order averaging, include more harmonics, or switch to non-averaged methods (Amato et al., 2018). Near tesseral or third-body resonances, double averaging is required, and highly eccentric or high-altitude orbits exhibit large long-periodic terms and large lunar perturbations that increase sensitivity to truncation choices.

In the elasticity solver, the method requires a star-shaped domain, piecewise-constant material coefficients in the angular sectors, and solution of a generalized eigenproblem whose cost can dominate when the angular discretization is large (Zhu et al., 18 Jun 2025). In the Grad–Shafranov solver, linearity depends on specific profile assumptions for pressure and current functions (Ciro et al., 2014). In the Noh-problem solver, the classic two-region structure imposes a geometry-specific restriction: in cylindrical and spherical symmetry one must have $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$8 and $$\dot\delta=\omega, \qquad M\dot\omega + D\omega = P_m - P_e(\delta).$$9 to retain the classic Noh solution structure (Gerberding et al., 20 Sep 2025).

A second misconception is that “semi-analytical” implies “more accurate than numerical integration” in a universal sense. The cited orbit-propagation paper directly contradicts that generalization: semi-analytical solvers may be faster in some regimes, but non-averaged regularized formulations can match or exceed both speed and accuracy in others (Amato et al., 2018). Conversely, a purely numerical baseline is not always the computational optimum when closed-form local propagation, averaging, or coordinatewise analytic resolution is available [(Duan et al., 2014); (Xu et al., 2022); (Zhu et al., 18 Jun 2025)].

The broader significance of SAS methods lies in this trade-off architecture. They preserve enough analytical structure to reduce repeated solves, exploit parallelism, enlarge time or space cells, or encode singular behavior in basis functions, while accepting locality, truncation, or model specialization as the price of efficiency. This suggests that SAS is less a single solver family than a recurrent design principle: isolate a tractable substructure, make it explicit, and use that explicit representation to reorganize the dominant computational cost.