AI Paper Atlas

Research map / Architectures and efficiency / Neural architecture search

Cool Graphs: Active Property Search Towards Quantum Nano-Refrigerators

Sofia Blyufer, Amir Kleiner, Uri Peskin

NeurIPS 2026 · Architectures and efficiency › Neural architecture search · approach: Differentiable architecture search (DARTS) · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes an active property search approach to efficiently find graph/nanodevice configurations that act as quantum refrigerators, avoiding expensive quantum-chemical computations for each candidate.

Problem

Finding nanodevice configurations that function as quantum refrigerators is a needle-in-a-haystack search: the many-body state space grows exponentially with device size and each candidate needs an expensive quantum-chemical calculation, with only 2-9% of random configurations showing the desired be

Approach

Treats configuration discovery as an active search problem over graphs, using sequential selection of candidates to find rare refrigerating configurations with far fewer expensive quantum-chemical evaluations than random sampling. Details beyond the abstract are not available.

Key results

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use uncertainty-driven (MC dropout) acquisition instead of random acquisitionGVAE active learning loop · 11 topologies, 16,000 oracle callspositives found: 670 -> 4477Improved
use deep-ensemble uncertainty instead of MC dropoutactive learning loop · 16,000 oracle callspositives found: 4477 -> 4925Improved
use uncertainty-driven acquisition instead of GP Bayesian optimizationsearch strategy · 16,000 oracle callspositives found: 1.46-1.53x more positivesImproved
retrain classifier on pool from uncertainty acquisition instead of random acquisitionGVAE · extrapolation testROC-AUC: 0.911 -> 0.934Improved
retrain classifier on pool from GP-BO acquisition (more positives) instead of randomGVAE · interpolation testROC-AUC: 0.932 -> 0.924 / 0.923No clear effect
add variational (KL) term vs plain autoencoder (ablation)graph autoencoder · interpolation testPR-AUC: 0.354 -> 0.409Improved
add auxiliary decoders (reconstruction) to discriminative classifiers (ablation)GAT/GCN classifiers · interpolation testPR-AUC: 0.07-0.12 higherImproved
remove the KL term entirely (ablation)GVAE · interpolation/extrapolation testPR-AUC: -0.017 (0.010)No clear effect
multiply KL weight by 10 (ablation)GVAE · interpolation testPR-AUC: -0.108Hurt
weaken reconstruction weight tenfold (ablation)GVAE · interpolation testPR-AUC: -0.083Hurt
raise latent dimension from 32 to 128 (ablation)GVAE · interpolation testPR-AUC: -0.138Hurt
inject random selection every eleventh active-learning cycle (ablation)uncertainty acquisition · per-cycle positive ratepositive rate: 0.280 -> 0.281No clear effect

Limitations

Experimental setup

Baselines: random sampling

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy