Energy Aligning: Concepts & Applications

• Energy aligning is a multidisciplinary term describing procedures that adjust energy parameters—such as logits, molecular levels, or alignment geometry—to meet desired operational criteria.
• Methods range from post-hoc free-energy calibration in biased classifiers and Boltzmann-weighted reweighting in generative models to quantum-level tuning at molecule–metal interfaces and mode-matching in resonant beam systems.
• Practical implementations demonstrate marked improvements including enhanced classification accuracies (e.g., 65.9% to 69.8% top-1 gains), efficient energy screening, and robust wireless power transfer via precise alignment.

Energy aligning is a non-unified term used across several research areas to denote procedures that correct an energy-related mismatch between a learned model, a physical interface, or an energy-transfer apparatus and a target operating condition. In recent literature it refers to post-hoc equalization of class free energies in biased classifiers, exact or Boltzmann-weighted reweighting of generative models, electronic energy level placement at molecule–metal interfaces, and self-alignment or mode-matching in resonant beam power-transfer systems (Zhao et al., 2021, Gu et al., 2024, Liu et al., 2019, Xiong et al., 8 May 2025). The common element is not a single algorithm but the deliberate manipulation of energies, logits, energy levels, or alignment geometry so that inference, sampling, transport, or transfer matches the specified criterion in each work.

1. Meanings and scope

In the cited literature, energy aligning appears in at least four distinct senses. Some works treat logits as negative energies and seek equal free energies across classes. Some treat reward or physical energy as a density-ratio signal that reweights a generative distribution. Some study the placement of molecular frontier levels relative to a substrate Fermi level. Others use “aligning” in the geometric sense: resonators, coils, or surface structures are arranged so that power transfer, communication, or anchoring remains effective across perturbations.

Domain	Object being aligned	Representative mechanism
Biased classifiers	Free energies of categories	Add calculated shift scalars onto the output logits during inference
Generative models	Model distribution to reward or physical energy	Exact Energy Preference Optimization, Boltzmann weights, energy/force rewards
Molecule–metal interfaces	Frontier orbital energies relative to $E_F$	$GW$ substrate screening or OT-RSH tuning
Power transfer and surfaces	Optical mode-matching, coil position, anchoring energy	Retroreflectors, MLE from RFID phase, light-controlled hybrid aligning layers

A recurring misconception is to treat these uses as interchangeable. The papers instead define different mathematical objects: $E_\theta(x,y)$ in an energy-based classifier, $r(x,c)$ or $E_{\rm physical}(x;c)$ in generative alignment, quasiparticle level shifts at interfaces, and geometric or anchoring conditions in optical and surface systems. This suggests that “energy aligning” is best understood as a family of domain-specific alignment procedures rather than a single standardized framework.

2. Free-energy equalization in biased classifiers

In "Energy Aligning for Biased Models" (Zhao et al., 2021), a discriminative classifier $f_\theta(x)\in\mathbb{R}^C$ is reinterpreted as an energy-based model by setting

$$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$

The class free energy is

$$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$

From the joint Gibbs distribution, the paper derives

$\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$

hence $-E_\theta(y=i)\propto \log p_\theta(y=i)$. Under a balanced target regime, equal class priors imply $GW$0 for all $GW$1. The central claim is that training on imbalanced data aligns free energies with the observed label distribution rather than the desired balanced one.

The method corrects this bias with additive shift scalars. For an anchor class $GW$2, each class $GW$3 receives

$GW$4

Since the integrals are intractable, they are approximated by Monte Carlo sampling: $GW$5 Inference then uses corrected logits

$GW$6

No network architecture change, no intervention in the standard learning paradigm, and no two-stage training are required. A clustering extension estimates one shift per frequency cluster and applies it to all logits in that cluster.

The reported evaluation covers class incremental learning and long-tailed recognition. On iNaturalist with ResNet-50 and 90 epochs, a baseline best decoupling result of approximately $GW$7 top-1 is improved to $GW$8. On ImageNet-LT with ResNeXt-50 and 90 epochs, a baseline best of approximately $GW$9 rises to $E_\theta(x,y)$0. On ImageNet100 in the 10-step class incremental setting, a baseline of approximately $E_\theta(x,y)$1 becomes $E_\theta(x,y)$2, and on ImageNet1000 in the corresponding 10-step setting, approximately $E_\theta(x,y)$3 becomes $E_\theta(x,y)$4. In this formulation, energy aligning is a post-hoc free-energy calibration procedure for class imbalance.

3. Energy-weighted alignment in generative modeling

In "Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization" (Gu et al., 2024), energy aligning is formulated as preference optimization for target-aware diffusion models. The paper contrasts standard DPO with Exact Energy Preference Optimization (E$E_\theta(x,y)$5PO), arguing that ordinary pairwise preference losses suffer from “winner-takes-all” overfitting. The aligned target distribution is

$E_\theta(x,y)$6

and the converged model distribution is derived in closed form as

$E_\theta(x,y)$7

The practical objective rescales diffusion-step preference terms by the exact Bradley–Terry probability $E_\theta(x,y)$8. On CrossDocked2020, the paper reports an Avg. Vina Score of $E_\theta(x,y)$9 for AliDiff-E$r(x,c)$0PO versus $r(x,c)$1 for the IPDiff baseline, with QED changing from $r(x,c)$2, SA from $r(x,c)$3, and diversity from $r(x,c)$4.

In "Aligning Protein Conformation Ensemble Generation with Physical Feedback" (Lu et al., 30 May 2025), Energy-based Alignment (EBA) fine-tunes a denoising diffusion model using physical energies from OpenMM 8.0 with CHARMM36+GBn2. The target Boltzmann distribution is

$r(x,c)$5

with mini-batch Boltzmann weights

$r(x,c)$6

The EBA loss is a weighted cross-entropy over model energies, and the implemented diffusion form uses a log-softmax over per-sample denoising losses rather than back-propagating through the physical energy oracle. The reported MD ensemble benchmark shows pairwise RMSD Pearson $r(x,c)$7 improving from $r(x,c)$8 to $r(x,c)$9, global RMSF Pearson $E_{\rm physical}(x;c)$0 from $E_{\rm physical}(x;c)$1 to $E_{\rm physical}(x;c)$2, and weak contacts Jaccard from $E_{\rm physical}(x;c)$3 to $E_{\rm physical}(x;c)$4, while runtime remains $E_{\rm physical}(x;c)$5 GPU s/sample.

Two later works extend the same general pattern. "Physio-DPO: Aligning LLMs with the Protein Energy Landscape to Eliminate Structural Hallucinations" (Meng, 2 Jan 2026) introduces a magnitude-aware weighting

$E_{\rm physical}(x;c)$6

inside a DPO objective, so that larger energy gaps drive larger updates. The paper reports self-consistency RMSD $E_{\rm physical}(x;c)$7 Å and foldability $E_{\rm physical}(x;c)$8 for Physio-DPO, compared with $E_{\rm physical}(x;c)$9 Å and $f_\theta(x)\in\mathbb{R}^C$0 for standard DPO. "Elign: Equivariant Diffusion Model Alignment from Foundational Machine Learning Force Fields" (Li et al., 29 Jan 2026) treats reverse diffusion as an MDP and fine-tunes the denoising policy with FED-GRPO using the terminal energy reward $f_\theta(x)\in\mathbb{R}^C$1, the terminal force reward $f_\theta(x)\in\mathbb{R}^C$2, and potential-based energy shaping on predicted clean geometries. On QM9, atom stability rises from $f_\theta(x)\in\mathbb{R}^C$3 to $f_\theta(x)\in\mathbb{R}^C$4, molecule stability from $f_\theta(x)\in\mathbb{R}^C$5 to $f_\theta(x)\in\mathbb{R}^C$6, and $f_\theta(x)\in\mathbb{R}^C$7 from $f_\theta(x)\in\mathbb{R}^C$8 to $f_\theta(x)\in\mathbb{R}^C$9; inference remains as fast as the unguided diffusion model because no energy evaluations are required during generation.

Taken together, these works define energy aligning as explicit density reweighting or reward shaping by exact reward differences, Boltzmann factors, or force-field surrogates. A plausible implication is that this strand of the literature uses “alignment” primarily in the distributional sense: the generator is pushed toward a preferred posterior while remaining anchored to a pretrained reference model.

4. Electronic energy level alignment at molecule–metal interfaces

At molecule–metal interfaces, energy aligning refers to the placement of molecular frontier levels relative to the metal Fermi level. "Energy Level Alignment at Molecule-Metal Interfaces from an Optimally-Tuned Range-Separated Hybrid Functional" (Liu et al., 2017) develops a self-consistent OT-RSH scheme in which the Coulomb operator is partitioned into short-range and long-range components, with exchange–correlation energy

$$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$0

For isolated molecules, $$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$1 is tuned by minimizing

$$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$2

At the interface, the image-charge energy

$$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$3

is used to retune $$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$4 so that

$$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$5

The reported results show level alignments agreeing to $$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$6 eV with UPS and inverse-UPS and work-function changes reproduced within $$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$7 eV of experiment across six interfaces. The method is explicitly self-consistent, unlike DFT+$$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$8, but its limitations include the use of a static classical image-charge term, a global scalar $$E_\theta(x,y=i)=-f_\theta(x)[i], \qquad p_\theta(y\mid x)=\frac{e^{-E_\theta(x,y)}}{\sum_{y'}e^{-E_\theta(x,y')}}.$$9, and the possibility of spurious gaps in metals.

Liu et al. address the same problem within $$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$0 in "Accelerating $$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$1-Based Energy Level Alignment Calculations for Molecule-Metal Interfaces Using a Substrate Screening Approach" (Liu et al., 2019). For weakly coupled interfaces, they introduce the additivity approximation

$$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$2

then compute the metal term in a smaller primitive cell and fold it into the interface supercell, while computing the molecular term in a reduced-$$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$3 cell and embedding it by real-space truncation. Because both components are genuine RPA polarizabilities, the method captures dynamical and nonlocal polarization without resorting to a classical image-charge formula or defining an image plane. For benzene/Al(111), the direct-$$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$4 $$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$5 step requires about $$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$6 CPU h and $$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$7 GB memory, whereas the substrate-screening procedure requires about $$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$8 CPU h and about $$E_\theta(y=i)=-\log\!\Bigl(\int_x e^{-E_\theta(x,i)}dx\Bigr) =-\log\!\Bigl(\int_x e^{f_\theta(x)[i]}dx\Bigr).$$9 GB, corresponding to $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$0 of the CPU cost and $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$1 of the memory. At a molecule–surface distance of $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$2 Å, the HOMO and LUMO differences relative to direct $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$3 are approximately $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$4 eV and $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$5 eV, and at $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$6 Å they drop below $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$7 eV.

In this interface literature, energy aligning denotes neither logit calibration nor geometric positioning. It denotes accurate prediction of quasiparticle level placement under screening, polarization, and hybridization.

5. Self-alignment and mode-matching in resonant beam SLIPT

"Duplex Self-Aligning Resonant Beam Communications and Power Transfer with Coupled Spatially Distributed Laser Resonator" (Xiong et al., 8 May 2025) uses “energy aligning” in the sense of robust self-alignment and mode-matching for simultaneous light information and power transfer (SLIPT). The coupled spatially distributed resonator (CSDR) consists of two sub-resonators sharing the partially reflective mirror $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$8: an intra-sub-resonator inside the transmitter and an extra-sub-resonator spanning the transmitter–receiver free-space link. Both $\log p_\theta(y=i)=-E_\theta(y=i)-\log Z_\theta,$9-$-E_\theta(y=i)\propto \log p_\theta(y=i)$0 and $-E_\theta(y=i)\propto \log p_\theta(y=i)$1-$-E_\theta(y=i)\propto \log p_\theta(y=i)$2 form a retroreflector pair, and $-E_\theta(y=i)\propto \log p_\theta(y=i)$3-$-E_\theta(y=i)\propto \log p_\theta(y=i)$4 and $-E_\theta(y=i)\propto \log p_\theta(y=i)$5-$-E_\theta(y=i)\propto \log p_\theta(y=i)$6 form a second retroreflector. Because a retroreflector returns any incoming beam parallel to its incident direction, the two cavities remain automatically aligned even as the free-space separation $-E_\theta(y=i)\propto \log p_\theta(y=i)$7 changes.

The key optical design choice is to place the $-E_\theta(y=i)\propto \log p_\theta(y=i)$8 waist exactly on the shared mirror $-E_\theta(y=i)\propto \log p_\theta(y=i)$9. With the single-pass ABCD matrix

$GW$00

the Gaussian-mode stability condition is

$GW$01

The fundamental Gaussian-beam parameter at $GW$02 is

$GW$03

and propagation obeys

$GW$04

The local beam radius is

$GW$05

The computed $GW$06 shows the unique waist always at $GW$07, independent of $GW$08, so the intra- and extra-cavities remain mode-matched without active pointing or tracking.

Power transfer is modeled with a Rigrod expression,

$GW$09

followed by

$GW$10

Mirror reflectivities affect both $GW$11 and $GW$12, so lowering $GW$13 can increase output power but also risk excessive cavity loss. The paper reports a numerical trade-off “ridge” in the $GW$14 plane and notes that one may choose $GW$15 and $GW$16 for approximately $GW$17 W charging with minimal free-space exposure.

A Type-I SHG crystal before $GW$18 generates a $GW$19 nm beam with single-pass efficiency

$GW$20

Because the second-harmonic beam is at half the wavelength of the resonant beam, it does not resonate and therefore does not interfere with the fundamental oscillation; the paper describes this spectral separation as an inherent suppression of echo interference even in a shared-path TDD scheme. Numerical illustrations show $GW$21 remaining in $GW$22 from a lower cut-off of approximately $GW$23 mm up to several metres, invariance of the beam radius at $GW$24 as $GW$25 sweeps from $GW$26 m to $GW$27 m, free-space extra-cavity power at approximately one-third of intra-cavity power, and symmetric duplex operation when $GW$28.

6. Alignment for power transfer and surface anchoring

In dynamic wireless power transfer, alignment is geometric and directly coupled to transfer efficiency. "Precise Coil Alignment for Dynamic Wireless Charging of Electric Vehicles with RFID Sensing" (Sun et al., 2023) models the round-trip RFID phase as

$GW$29

with

$GW$30

Assuming Gaussian noise after phase unwrapping, maximum-likelihood estimation is equivalent to minimizing the squared residuals over $GW$31. The implementation uses a coarse-to-fine grid search followed by local refinement such as Gauss–Newton iteration. Laboratory tests report a mean lateral error of approximately $GW$32 cm over $GW$33 runs, and field tests report a mean error of approximately $GW$34 cm over $GW$35 runs; the abstract characterizes the method as capable of achieving sub-10 cm accuracy. The practical significance is explicit: a $GW$36 cm lateral shift can reduce coupling by more than $GW$37, translating to a $GW$38–$GW$39 drop in end-to-end efficiency, whereas keeping $GW$40 m maintains $GW$41 and yields $GW$42 transfer efficiency even at $GW$43 kW-class power levels.

In liquid-crystal technology, alignment concerns anchoring energy rather than transport efficiency. "Light-controllable hybrid aligning layer based on LIPSS on sapphire surface and PVCN-F film" (Gvozdovskyy et al., 2023) studies hybrid aligning layers composed of laser-structured sapphire and a photoaligning PVCN-F coating. The azimuthal anchoring energy follows the Rapini–Papoular form

$GW$44

and in the twist-cell measurement geometry

$GW$45

For grooved surfaces, the Berreman estimate is

$GW$46

so groove depth $GW$47 and period $GW$48 directly influence anchoring. The reported non-irradiated hybrid layers show $GW$49. UV irradiation with polarization parallel to grooves decreases $GW$50 monotonically to approximately $GW$51, close to zero anchoring, whereas polarization perpendicular to grooves increases $GW$52 up to $GW$53. AFM shows groove depth changing from about $GW$54 nm after PVCN-F coating to about $GW$55 nm for $GW$56 grooves and about $GW$57 nm for $GW$58 grooves, consistent with the Berreman scaling. Contact-angle measurements show a direct inverse correlation: larger $GW$59 corresponds to smaller $GW$60, and vice versa.

These hardware and surface studies indicate that “aligning” can denote geometric positioning or controllable anchoring-energy engineering rather than probabilistic score correction. Across the broader literature, the term remains domain-bound: in one context it is a one-line post-hoc logit shift, in another a Boltzmann-weighted training objective, in another a quasiparticle screening calculation, and in another a resonator, coil, or surface design that preserves power transfer or orientation under perturbation.