---
title: Energy-Structured Causal Models (E-SCM)
url: https://www.emergentmind.com/topics/energy-structured-causal-models-e-scm
type: topic
---

# Energy-Structured Causal Models (E-SCM)

Searching arXiv for the specified papers and closely related work to ground the article.
Energy-Structured Causal Models (E‑SCMs) are causal models in which mechanisms are represented as constraints, expressed either as energy functions or vector fields, rather than as explicit input–output assignments. In this formulation, system states are equilibria—typically minima of a total energy or fixed points of dynamics—and interventions are implemented as local surgeries on the corresponding constraints, followed by re-equilibration [2510.22050]. Within this broad formulation, the demand-forecasting model in "Causal Inference in Energy Demand Prediction" functions as a domain-specific instance of an E‑SCM: variables, edges, and structural equations are organized around interpretable energy subsystems such as HVAC needs, routine activity needs, and lighting needs, and the resulting structure is used both for causal analysis and Bayesian prediction [2512.11653].

## 1. Formal definition and core semantics

A static E‑SCM is defined over endogenous latent variables $\mathbf Z=(Z_1,\dots,Z_n)$ and exogenous variables $\mathbf U=(U_1,\dots,U_n)$, together with a DAG $\mathcal G$ specifying parent sets $\mathrm{PA}(i)$, local mechanism energies $\{E_i\}_{i=1}^n$, exogenous energies $\{E_{U_i}\}_{i=1}^n$, and an optional global term $E_{\mathrm{global}}$ [2510.22050]. The local mechanisms satisfy
$$
E_i:\ \mathcal Z_i\times \mathcal Z_{\mathrm{PA}(i)} \times \mathcal U_i \to \mathbb R,
$$
while exogenous energies satisfy
$$
E_{U_i}:\ \mathcal U_i\to\mathbb R.
$$
The total energy is
$$
E(\mathbf z,\mathbf u) = \sum_{i=1}^n E_i\bigl(z_i\mid z_{\mathrm{PA}(i)},u_i\bigr) + \sum_{i=1}^n E_{U_i}(u_i) + E_{\mathrm{global}}(\mathbf z,\mathbf u).
$$
Admissible states are equilibria satisfying
$$
\nabla_{\mathbf z,\mathbf u} E(\mathbf z,\mathbf u) = 0,
$$
typically under convexity or uniqueness conditions ensuring a unique equilibrium for each exogenous context [2510.22050].

A dynamic E‑SCM replaces scalar energies with local vector fields,
$$
\frac{dz_i}{dt} = F_i\bigl(z_i, z_{\mathrm{PA}(i)}, u_i\bigr),\qquad i=1,\dots,n,
$$
so that mechanisms are differential constraints and interventions operate by rewriting selected local dynamics [2510.22050]. Static and dynamic forms coincide when the dynamics are gradient flows, $\dot z = -\nabla_z E$, but the dynamic formalism also admits non-gradient systems.

The central semantic shift is from explicit structural assignments to equilibrium selection under constraints. In a classical SCM, a mechanism computes $X_i$ from $X_{\mathrm{PA}(i)}$ and $U_i$. In an E‑SCM, the mechanism defines which joint configurations are preferred or admissible, and the realized state is determined by global equilibrium across all such constraints [2510.22050]. This makes the unit of intervention an energy term or vector-field component rather than a structural equation in explicit solved form.

## 2. Relation to classical SCMs

The formal relationship between E‑SCMs and standard SCMs is established through a reduction theorem. For an E‑SCM with energy
$$
E(x;u) = \sum_{i=1}^n \phi_i\bigl(x_i,x_{\mathrm{PA}(i)},u_i\bigr),
$$
under locality, blockwise strict convexity, global strict convexity and coercivity, modular interventions, exogenous stability, and measurability, one defines local best-response maps
$$
f_i(x_{\mathrm{PA}(i)},u_i) \;:=\; \operatorname*{arg\,min}_{\tilde x_i} \phi_i(\tilde x_i, x_{\mathrm{PA}(i)}, u_i).
$$
The unique global minimizer $x^\star(u)$ then satisfies
$$
x_i^\star(u) = f_i\bigl(x^\star_{\mathrm{PA}(i)}(u), u_i\bigr),
$$
and conversely any solution of
$$
X_i = f_i(X_{\mathrm{PA}(i)}, U_i)
$$
is the unique minimizer of the energy [2510.22050].

The induced SCM has structural equations
$$
X_i := f_i\bigl(X_{\mathrm{PA}(i)}, U_i\bigr)
$$
and the same exogenous law $P_U$. Under the theorem’s assumptions, the E‑SCM and induced SCM are observationally equivalent, interventionally equivalent, and counterfactually equivalent [2510.22050]. This means that E‑SCMs do not discard SCM semantics; rather, they reparameterize those semantics in a constraint-based form.

This equivalence is important for causal inference because it preserves the standard abduction–intervention–prediction interpretation. In E‑SCM form, abduction corresponds to constrained energy minimization to infer exogenous variables and latent states; intervention corresponds to editing local energies; prediction corresponds to re-solving for equilibrium in the edited system [2510.22050]. A plausible implication is that E‑SCMs are best understood not as a rejection of SCMs, but as an alternative mechanism representation particularly suited to latent-variable and equilibrium-based settings.

## 3. Mechanisms, interventions, and counterfactuals

In E‑SCMs, mechanisms compose additively:
$$
E(\mathbf z,\mathbf u) = \sum_i E_i(z_i \mid z_{\mathrm{PA}(i)},u_i) + \sum_i E_{U_i}(u_i) + E_{\mathrm{global}}(\mathbf z,\mathbf u).
$$
Because mechanisms are separate energy terms, adding, removing, or editing a mechanism is literally addition, deletion, or replacement of terms in the total energy [2510.22050]. This additive structure is the energy analogue of modular decomposition in SCMs.

Hard interventions $\mathrm{do}(Z_j=z_j^\ast)$ are implemented by deleting the original local mechanism for $Z_j$ and imposing an infinite barrier outside the intervened value:
$$
E^{\mathrm{do}(Z_j=z_j^\ast)}(\mathbf z,\mathbf u) = \begin{cases}
\displaystyle \sum_{i\neq j} E_i(z_i\mid z_{\mathrm{PA}(i)},u_i) + \sum_{i=1}^n E_{U_i}(u_i) + E_{\mathrm{global}}(\mathbf z,\mathbf u), & \text{if } z_j = z_j^\ast,\\[0.5em]
+\infty, & \text{otherwise.}
\end{cases}
$$
Soft interventions replace the local term by a convex combination of original and intervened energies,
$$
\widetilde{E}_j(z_j\mid z_{\mathrm{PA}(j)},u_j) = (1-\lambda) E_j^{\text{original}}(z_j\mid z_{\mathrm{PA}(j)},u_j) + \lambda\, E_j^{\text{intervene}}(z_j\mid z_{\mathrm{PA}(j)},u_j),
$$
with $\lambda\in[0,1]$ [2510.22050]. The framework also includes disjunctive interventions $do(Z_j\in\mathcal S)$ either as policy-free families of singleton interventions or by introducing a control energy over admissible values.

Counterfactuals follow the standard three-step structure. First, abduction solves
$$
(\hat{\mathbf z},\hat{\mathbf u})
=
\arg\min_{\mathbf z,\mathbf u} E(\mathbf z,\mathbf u)
\quad\text{s.t.}\quad
z_k = z_k^{\text{cond}} \ \forall k\in\mathcal O.
$$
Second, intervention edits the relevant local energies while holding the abducted exogenous variables fixed. Third, prediction re-minimizes the edited energy to obtain post-intervention outcomes [2510.22050]. The formal identity with SCM counterfactual semantics holds under the reduction theorem’s assumptions.

## 4. Structural principles: LAP, ICM, and gauge ambiguity

A distinctive feature of the E‑SCM program is that structural principles are expressed as differential conditions on energies or vector fields. The Locality-Autonomy Principle (LAP) requires that a node’s mechanism not depend on non-descendants, either through their states or through their parameters. For a static E‑SCM, if $\widetilde{E}_i^{(A)}$ denotes the effective energy for node $i$ relative to a module $A$, then for non-descendant $i\notin\mathrm{Desc}(A)$ the required conditions are
$$
\frac{\partial^2 \widetilde{E}_i^{(A)}}{\partial z_i\,\partial z_A} = 0, \qquad \frac{\partial^2 \widetilde{E}_i^{(A)}}{\partial z_i\,\partial \theta_A} = 0.
$$
Training can penalize violations through
$$
\mathcal L_{\text{LAP}}^{\text{static}}
= \sum_{A}\sum_{i\in\mathrm{NonDesc}(A)} \lambda_{A,i}\, \mathbb E_{\mathcal Q} \left[
  \Bigl\|
    \frac{\partial^2 \widetilde{E}_i^{(A)}}{\partial z_i \partial z_A}
  \Bigr\|^2 \right]
+ \mu_{A,i}\, \mathbb E_{\mathcal Q} \left[
  \Bigl\|
    \frac{\partial^2 \widetilde{E}_i^{(A)}}{\partial z_i \partial \theta_A}
  \Bigr\|^2 \right]
$$
[2510.22050].

Independent Causal Mechanisms (ICM) is formulated in parameter space. Defining the residual
$$
G_i(z,u;\theta) := \frac{\partial E(z,u;\theta)}{\partial z_i},
$$
the conditions are
$$
\frac{\partial G_i}{\partial \theta_{\mathrm{PA}(i)}} = 0, \qquad
\frac{\partial^2 G_i}{\partial \theta_{\mathrm{PA}(i)}\,\partial \theta_i} = 0.
$$
Corresponding penalties are introduced to enforce approximate ICM during learning [2510.22050]. CAP, the Compositional Autonomy Principle, is treated as an outlook principle concerning transport of mechanism templates and preservation of intervention responses across domains.

The E‑SCM literature also uses a gauge-theoretic analysis of encoder–energy pairs to explain why empirical risk minimization can yield causally opaque internal representations. If $f:\mathcal X\to\mathcal Z$ is an encoder and $E(z)$ an energy on the latent space, then gauge transformations of the form
$$
\gamma:(f,E)\mapsto (\gamma_f\circ f, \; E\circ\gamma_z^{-1} + c)
$$
can preserve observational heads while altering intervention-relevant internal structure [2510.22050]. The paper distinguishes the larger set of transformations preserving observational heads from the smaller set preserving causal heads, and interprets the difference as the locus of fractured or entangled representations. Probes such as absolute per-module energies, gradients, energy differences, and Hessians progressively shrink this observational gauge freedom.

A plausible implication is that E‑SCMs attempt to make causal modularity directly inspectable and regularizable at the representation level, not solely at the level of externally observed statistical relationships.

## 5. Energy demand prediction as a worked example

"Causal Inference in Energy Demand Prediction" presents a structural causal model for load forecasting that can be read as a concrete E‑SCM for an energy system [2512.11653]. The observed exogenous root variables are calendar variables hour of day $H\in\{0,\dots,23\}$ and month of year $M\in\{1,\dots,12\}$. Observed weather variables are temperature $T$, relative humidity $R.H.$, wind speed $W$, and solar radiation $\text{Rad}$. Latent or conceptual variables are HVAC needs, routine activity needs, and lighting needs. The target is energy demand $E$.

The proposed DAG organizes these variables into physically and behaviorally meaningful subsystems. Calendar variables affect weather; weather affects HVAC needs; calendar variables affect routine activities and hence routine activity needs; calendar variables and solar radiation affect lighting needs; and the three need-based subsystems combine to determine total demand [2512.11653]. In conceptual structural-equation form,
$$
\begin{aligned}
H &= f_H(U_H) \\
M &= f_M(U_M) \\
T &= f_T(H, M, U_T) \\
R.H. &= f_{R.H.}(H, M, U_{R.H.}) \\
W &= f_W(H, M, U_W) \\
\text{Rad} &= f_{\text{Rad}}(H, M, U_{\text{Rad}}) \\
\text{HVAC} &= f_{\text{HVAC}}(T, R.H., W, U_{\text{HVAC}}) \\
\text{activity} &= f_{\text{act}}(H, M, U_{\text{act}}) \\
\text{lighting} &= f_{\text{light}}(H, M, \text{Rad}, U_{\text{light}}) \\
E &= f_E(\text{HVAC}, \text{activity}, \text{lighting}, U_E).
\end{aligned}
$$
The joint distribution factorizes as
$$
p(H,M,T,R.H.,W,\text{Rad},E) = p(H)p(M)p(T\mid H,M)p(R.H.\mid H,M)p(W\mid H,M)p(\text{Rad}\mid H,M)p(E\mid T,R.H.,W,H,M,\text{Rad}).
$$

Temperature is modeled conditionally on hour and month through a harmonic series:
$$
\begin{aligned}
T &= \underbrace{\sum_{j=1}^{n} \left(c_j \sin\left(\frac{2\pi j M}{12}\right) + c_j^{*} \cos\left(\frac{2\pi j M}{12}\right)\right)}_{\text{yearly seasons}} \\
&\quad + \underbrace{\sum_{j=1}^{n} \left(d_j \sin\left(\frac{2\pi j H}{24}\right) + d_j^{*} \cos\left(\frac{2\pi j H}{24}\right)\right)}_{\text{daily temperature cycle}} \\
&\quad + T_{\text{base}} + e_{\text{temp}},
\end{aligned}
$$
with Gaussian noise [2512.11653]. Relative humidity, solar radiation, and wind speed are assigned Gaussian priors in the Bayesian implementation; the paper explicitly treats this as a modeling simplification rather than a causal claim of independence from calendar variables.

The temperature response of demand is represented through a V-shape transform
$$
f(T) = |T - T_{\text{mid}}|,
$$
with $T_{\text{mid}} = 56^\circ\mathrm{F}$ [2512.11653]. Base HVAC demand is then
$$
E_{\text{base}} = k f(T) + E_0.
$$
Humidity contributes conditionally in the hot regime:
$$
E_{\text{humid}} = R.H.\,\delta(T > T_{R.H.}),
$$
where $T_{R.H.}$ is chosen as $70^\circ\mathrm{F}$ by grid search maximizing correlation with demand. Wind contributes asymmetrically across thermal regimes:
$$
E_{\text{wind}} = W\,\delta(T < T_{w,1}) - \lambda W\,\delta(T > T_{w,2}).
$$
Daily activity demand is modeled as an hourly harmonic term,
$$
E_{\text{daily}} = \sum_{j=1}^{n} \left(a_j \sin\left(\frac{2\pi j H}{24}\right) + a_j^{*} \cos\left(\frac{2\pi j H}{24}\right)\right),
$$
and yearly activity demand as a monthly harmonic term,
$$
E_{\text{yearly}} = \sum_{j=1}^{n} \left(\alpha_j \sin\left(\frac{2\pi j M}{12}\right) + \alpha_j^{*} \cos\left(\frac{2\pi j M}{12}\right)\right).
$$
Lighting demand is
$$
E_{\text{light}} = \exp(-\beta \cdot \text{Rad}) \,\delta(\text{active hour}),
$$
with the active-hour indicator equal to 1 between 5 a.m. and 12 a.m. and 0 otherwise [2512.11653].

Total demand is decomposed additively as
$$
E = \underbrace{E_{\text{base}} + E_{\text{humid}} + E_{\text{wind}}}_{\text{HVAC needs}}
+ \underbrace{E_{\text{light}}}_{\text{lighting needs}}
+ \underbrace{E_{\text{daily}} + E_{\text{yearly}}}_{\text{activity consumption}}.
$$
This additive subsystem structure is the sense in which the model is "energy-structured" in domain terms: functions are aligned to interpretable physical and behavioral mechanisms rather than to a purely correlational predictor [2512.11653].

## 6. Causal identification, Bayesian implementation, and empirical behavior

The demand paper emphasizes confounding explicitly. For humidity, the DAG contains the paths
$$
H \to R.H., \qquad H \to \text{routine activities} \to E, \qquad R.H. \to \text{HVAC needs} \to E.
$$
Thus $H$ is a confounder of humidity and energy demand in Pearl’s sense [2512.11653]. A naive regression yields
$$
\text{corr}(R.H., E) = -0.07,
$$
which is inconsistent with the domain expectation that higher humidity increases cooling demand. Conditional on hour and temperature, the sign becomes positive in hot regimes. The paper interprets this as a positive direct causal path from humidity to demand together with a negative backdoor path through hour.

For temperature, the target causal quantity is $E[y\mid do(x)]$ with $x=T$, $y=E$, and $z=H$. The paper states that $H$ satisfies the backdoor criterion for $(T,E)$, so
$$
P(y\mid do(x)) = \sum_z P(y\mid x,z)P(z).
$$
If
$$
E[y\mid x,z] = a' z + b' x^* + c',
$$
with $x^* = |T-T_{\text{mid}}|$, then
$$
E[y\mid do(x)] = a' E[z] + b' x^* + c'.
$$
Accordingly, the adjusted regression coefficient $b'$ coincides with the causal coefficient in
$$
E[y \mid do(x)] = b x^* + c.
$$
The paper contrasts this causal adjustment with a non-causal approach that regresses on temperature and then separately on the daily cycle; the latter is reported to produce biased temperature coefficients [2512.11653].

The Bayesian model operationalizes the causal structure by placing priors over mechanism parameters and using stochastic variational inference in Pyro with an AutoNormal guide [2512.11653]. For each hour-indexed datum,
$$
E_t \sim \mathcal{N}\big(E_{\text{base},t} + E_{\text{humid},t} + E_{\text{wind},t} + E_{\text{light},t} + E_{\text{daily},t} + E_{\text{yearly},t}, \sigma_E^2\big).
$$
The paper provides examples of Gaussian priors such as
$$
k \sim \mathcal{N}(20.0,\sigma_2^2), \qquad
E_0 \sim \mathcal{N}(3485.0,\sigma_2^2), \qquad
a_1 \sim \mathcal{N}(-150.0,\sigma_2^2), \qquad
a_1^* \sim \mathcal{N}(136.0,\sigma_2^2),
$$
with $\sigma_2 = 40$.

Performance is evaluated using Mean Absolute Percentage Error,
$$
\text{MAPE} = \frac{100\%}{n} \sum_{t=1}^{n} \left|\frac{E_t - \hat E_t}{E_t}\right|.
$$
The reported results are train MAPE $3.23\%$, test MAPE $3.84\%$, and 5-fold cross-validation average MAPE $3.88\%$ across two years of data [2512.11653]. The paper also reports that ignoring confounding yields 47.8% deviation in temperature coefficients and 12.5% worse out-of-sample MAPE than the causal approach.

The model’s seasonal analysis attributes lower winter variance in demand to decoupling between temperature-driven HVAC fluctuations and activity-driven fluctuations. In summer, midday heat and midday activity are aligned, producing positive covariance between subsystem peaks; in winter, heating demand peaks in morning and evening while activity remains concentrated around midday, so the covariance is lower [2512.11653]. The paper treats this heteroscedasticity as emergent from the causal structure rather than as a separately fitted variance model.

## 7. Position in causal machine learning and generalization across energy systems

E‑SCMs occupy an intermediate position between SCMs, energy-based models, dynamical systems, and domain-structured Bayesian models. Relative to SCMs, they are more general in representation but, under mild assumptions, equivalent in causal semantics [2510.22050]. Relative to graphical models, they factor a deterministic energy rather than a probability distribution. Relative to ordinary energy-based models, they add a DAG, explicit intervention semantics, and diagnostic principles such as LAP and ICM. Relative to unconstrained deep learning, they aim to provide mechanism-level editability and causal interpretability rather than observational adequacy alone [2510.22050].

The foundational paper is mainly conceptual and does not present full empirical case studies, but it sketches an architectural ecosystem comprising adaptors, mechanisms, actuators, and probes. Adaptors map raw neural representations into causal latents and provide initializers for equilibrium solves. Mechanisms are parametric energies or vector fields with DAG-based masking and LAP/ICM penalties. Actuators perform intervention by editing specific local energies. Probes expose internal quantities such as equilibria, energies, gradients, and Hessians for diagnostics and evaluation [2510.22050]. Training combines a primary observational objective with structural regularizers.

The demand-prediction paper suggests how this outlook can be specialized to energy applications. Its construction pipeline proceeds from domain knowledge to a causal DAG, from DAG to structural equations, and from structural equations to a Bayesian predictive model [2512.11653]. The same document proposes broader construction steps for other energy domains: define physical and behavioral subsystems; list exogenous drivers, latent process states, and outputs; draw the causal DAG; specify mechanism-reflective structural equations with piecewise or nonlinear effects; encode the model as a Bayesian program; and validate causal structure through conditional independencies, backdoor reasoning, and causal-versus-non-causal comparisons.

Examples mentioned include renewable generation, industrial loads, and EV charging [2512.11653]. In these applications, the E‑SCM perspective would retain the same organizing principle: variables and functional forms are arranged around physically and institutionally meaningful mechanisms, interventions are evaluated as edits to those mechanisms, and robustness under non-stationarity is sought through structural rather than purely correlational inductive bias. This suggests that the term "Energy-Structured Causal Model" has two closely related uses in the current literature: a general causal formalism based on energies and equilibria [2510.22050], and a domain-specific modeling stance in which causal structure is organized around interpretable subsystems of an energy application [2512.11653].

Source: https://www.emergentmind.com/topics/energy-structured-causal-models-e-scm