--- title: 'Energy Aligning: Concepts & Applications' url: https://www.emergentmind.com/topics/energy-aligning type: topic --- # Energy Aligning: Concepts & Applications 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 [2106.03343] [2407.01648] [1904.02012] [2505.05107]. 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" [2106.03343], 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 \(E_\theta(y=i)=E_\theta(y=j)\) for all \(i,j\). 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 \(i\), each class \(j\) receives \[ \Delta_j = \log\frac{\int_x e^{f_\theta(x)[i]}dx}{\int_x e^{f_\theta(x)[j]}dx}. \] Since the integrals are intractable, they are approximated by Monte Carlo sampling: \[ \widehat\Delta_j = \log\sum_s e^{f_\theta(x_s)[i]} - \log\sum_s e^{f_\theta(x_s)[j]}. \] Inference then uses corrected logits \[ f'_\theta(x)[j]=f_\theta(x)[j]+\Delta_j. \] 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 \(65.9\%\) top-1 is improved to \(69.8\%\). On ImageNet-LT with ResNeXt-50 and 90 epochs, a baseline best of approximately \(51.8\%\) rises to \(52.7\%\). On ImageNet100 in the 10-step class incremental setting, a baseline of approximately \(90.2\%/68.8\%\) becomes \(91.2\%/71.1\%\), and on ImageNet1000 in the corresponding 10-step setting, approximately \(85.7\%/64.1\%\) becomes \(87.0\%/65.7\%\). 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" [2407.01648], energy aligning is formulated as preference optimization for target-aware diffusion models. The paper contrasts standard DPO with Exact Energy Preference Optimization (E\(^2\)PO), arguing that ordinary pairwise preference losses suffer from “winner-takes-all” overfitting. The aligned target distribution is \[ p^*(x\mid c)\propto p_{\rm ref}(x\mid c)\,e^{r(x,c)}, \] and the converged model distribution is derived in closed form as \[ p_\theta(x\mid c)\propto p_{\rm ref}(x\mid c)\cdot \exp\!\bigl[(1/\beta)\,r(x,c)\bigr]. \] The practical objective rescales diffusion-step preference terms by the exact Bradley–Terry probability \(\sigma(r^w-r^l)\). On CrossDocked2020, the paper reports an Avg. Vina Score of \(-7.07\) for AliDiff-E\(^2\)PO versus \(-6.42\) for the IPDiff baseline, with QED changing from \(0.52\to0.50\), SA from \(0.61\to0.57\), and diversity from \(0.74\to0.73\). In "Aligning Protein Conformation Ensemble Generation with Physical Feedback" [2505.24203], 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 \[ p_B(x\mid c)=\frac{e^{-\beta E_{\rm physical}(x;c)}}{Z}, \] with mini-batch Boltzmann weights \[ w^i=\frac{e^{-\beta E_{\rm physical}(x^i)}}{\sum_j e^{-\beta E_{\rm physical}(x^j)}}. \] 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 \(r\) improving from \(0.43\) to \(0.62\), global RMSF Pearson \(r\) from \(0.50\) to \(0.71\), and weak contacts Jaccard from \(0.23\) to \(0.65\), while runtime remains \(0.9\) GPU s/sample. Two later works extend the same general pattern. "Physio-DPO: Aligning Large Language Models with the Protein Energy Landscape to Eliminate Structural Hallucinations" [2601.00647] introduces a magnitude-aware weighting \[ \Psi(\Delta E)=\lambda_{\max}\,\sigma\!\Bigl(\frac{\Delta E-\mu}{\tau}\Bigr) \] inside a DPO objective, so that larger energy gaps drive larger updates. The paper reports self-consistency RMSD \(1.28\) Å and foldability \(92.8\%\) for Physio-DPO, compared with \(1.82\) Å and \(83.6\%\) for standard DPO. "Elign: Equivariant Diffusion Model Alignment from Foundational Machine Learning Force Fields" [2601.21985] treats reverse diffusion as an MDP and fine-tunes the denoising policy with FED-GRPO using the terminal energy reward \(r_0^{(E)}(z_0)=-E_\phi(z_0)\), the terminal force reward \(r_0^{(F)}(z_0)=-\|\mathbf F_\phi(z_0)\|_{\rm rms}^2\), and potential-based energy shaping on predicted clean geometries. On QM9, atom stability rises from \(98.70\%\) to \(99.33\%\), molecule stability from \(82.00\%\) to \(93.70\%\), and \(V\times U\) from \(90.70\%\) to \(95.31\%\); 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" [1703.02004] 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 \[ E_{\rm XC}[\rho] = \alpha E_{\rm X,SR}^{\rm EXX} +(1-\alpha)E_{\rm X,SR}^{\rm GGA} +(\alpha+\beta)E_{\rm X,LR}^{\rm EXX} +(1-\alpha-\beta)E_{\rm X,LR}^{\rm GGA} +E_{\rm C}^{\rm GGA}. \] For isolated molecules, \(\mu\) is tuned by minimizing \[ J(\mu)= \bigl|\varepsilon_{\rm HOMO}^\mu(N)+{\rm IP}^\mu(N)\bigr|^2 + \bigl|\varepsilon_{\rm HOMO}^\mu(N+1)+{\rm IP}^\mu(N+1)\bigr|^2. \] At the interface, the image-charge energy \[ P=\frac{1}{4(z-z_0)} \] is used to retune \(\beta\) so that \[ \varepsilon_{\rm HOMO}^{\mu}(\beta)-\varepsilon_{\rm HOMO}^{\mu}(\beta_0)=P. \] The reported results show level alignments agreeing to \(\le 0.2\) eV with UPS and inverse-UPS and work-function changes reproduced within \(\sim 0.1\) eV of experiment across six interfaces. The method is explicitly self-consistent, unlike DFT+\(\Sigma\), but its limitations include the use of a static classical image-charge term, a global scalar \(\beta\), and the possibility of spurious gaps in metals. Liu et al. address the same problem within \(GW\) in "Accelerating \(GW\)-Based Energy Level Alignment Calculations for Molecule-Metal Interfaces Using a Substrate Screening Approach" [1904.02012]. For weakly coupled interfaces, they introduce the additivity approximation \[ \chi^0_{\rm tot}(\mathbf r,\mathbf r') \approx \chi^0_{\rm mol}(\mathbf r,\mathbf r') + \chi^0_{\rm metal}(\mathbf r,\mathbf r'), \] 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-\(z\) 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-\(GW\) \(\chi^0\) step requires about \(3550\) CPU h and \(1257\) GB memory, whereas the substrate-screening procedure requires about \(1016\) CPU h and about \(170\) GB, corresponding to \(\sim 30\%\) of the CPU cost and \(\sim 14\%\) of the memory. At a molecule–surface distance of \(3.24\) Å, the HOMO and LUMO differences relative to direct \(GW\) are approximately \(0.16\) eV and \(0.05\) eV, and at \(4.24\) Å they drop below \(0.1\) 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" [2505.05107] 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 \(M_2\): an intra-sub-resonator inside the transmitter and an extra-sub-resonator spanning the transmitter–receiver free-space link. Both \(M_1\)-\(L_1\) and \(L_2\)-\(M_2\) form a retroreflector pair, and \(L_3\)-\(M_2\) and \(L_4\)-\(M_3\) 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 \(d_t\) changes. The key optical design choice is to place the \({\rm TEM}_{00}\) waist exactly on the shared mirror \(M_2\). With the single-pass ABCD matrix \[ \mathbf M_r= \begin{pmatrix} A_r & B_r\\ C_r & D_r \end{pmatrix}, \qquad g_1=A_r,\; g_2=D_r,\; L_{\rm eq}=B_r, \] the Gaussian-mode stability condition is \[ 00.9\) and yields \(>90\%\) transfer efficiency even at \(100\) 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" [2303.01749] studies hybrid aligning layers composed of laser-structured sapphire and a photoaligning PVCN-F coating. The azimuthal anchoring energy follows the Rapini–Papoular form \[ F_a(\theta)=\frac12 W\sin^2(\theta-\theta_0), \] and in the twist-cell measurement geometry \[ W=\frac{2K_{22}\varphi}{d\,\sin 2(\varphi_0-\varphi)}. \] For grooved surfaces, the Berreman estimate is \[ W_B=\frac{\pi^2}{2}\,\frac{K A^2}{\Lambda^3}, \] so groove depth \(A\) and period \(\Lambda\) directly influence anchoring. The reported non-irradiated hybrid layers show \(W\approx(0.76\ldots1.1)\times10^{-6}\,{\rm J\,m^{-2}}\). UV irradiation with polarization parallel to grooves decreases \(W\) monotonically to approximately \((2\ldots3)\times10^{-8}\,{\rm J\,m^{-2}}\), close to zero anchoring, whereas polarization perpendicular to grooves increases \(W\) up to \(19.7\times10^{-6}\,{\rm J\,m^{-2}}\). AFM shows groove depth changing from about \(50\) nm after PVCN-F coating to about \(38\) nm for \(E\parallel\) grooves and about \(62\) nm for \(E\perp\) grooves, consistent with the Berreman scaling. Contact-angle measurements show a direct inverse correlation: larger \(W\) corresponds to smaller \(\beta\), 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. Source: https://www.emergentmind.com/topics/energy-aligning