~/aixsci
weekly updates

$ aixsci --about

How AI is actually being used in science.

One record per paper: what the AI did, how the work was checked, and what the paper left out. Checked against the source, checked by a person, never silently changed.

3. Train and select occupancy models4. Predict sublattice and Fe-site occupancies8. Prospective screening of rare-earth dopants4. Train CNN classifiers per data product5. Classify literature candidates with the SED ensemble vote6. Evaluate on held-out test set and external contaminant catalogues2. Predict 2-mer and N-mer structures with AlphaFold4. Train autoencoder to compress fingerprints to two dimensions5. Cluster fingerprints into ordered and disordered classes6. Train MLP to carry the clustering boundary to unseen configurations7. Label all monomers at every step of the cooling cycle3. Pre-train on inverse protein folding4. Transfer-learn affinity predictor5. Predict ln(KD) and ΔΔG for mutants6. Read residue-level attention scores7. Screen multi-site mutation combinations3. Train binary CNN classifier4. Test detector on synthetic and real profiles7. Run detector over selected time profiles3. Predict protein structures with ESMFold4. Encode protein sequence–structure representation5. Train base classifiers and NAM meta-learner6. Evaluate on independent test sets and compare with published values7. Screen the E. coli O157:H7 proteome for candidates8. Attribute predictions with SHAP and sequence masking3. Fit and tune DT, RF and PINN wear models4. Predict wear on held-out data6. Attribute wear to input parameters with SHAP4. Fit and compare regression algorithms5. Predict DoR for the independent sample7. Measure feature importance across DoR thresholds5. Sample posterior of initial field and bias parameters5. Fit force constants by sparse regression6. Test force prediction on independent ensemble3. Train DeepVel on simulated granulation4. Compare DeepVel and FLCT against simulated flows5. Infer transverse velocities from IMaX intensitygrams6. Train the GA-initialised BP neural network7. Evaluate the trained network on the withheld set8. Run the model for sensitivity levels and kinetic curves4. Train scaffold model for reference period5. Train full model with hybrid objective and conformal calibration6. Rolling-window period inference5. Train element-type-specific Hamiltonian networks on dimers6. Predict aggregate Hamiltonians and excited-state properties5. Train CNN mass estimators6. Predict masses on held-out folds7. Saliency-based interpretability study4. Train artefact-rejection classifier iteratively5. Score all candidates and queue the highest-ranked4. Train and test separate SMC and LMC classifiers5. Classify all VMC sources6. Compute and inspect feature importances5. Train six classifier families under star-disjoint splits6. Score held-out cadences and apply decision threshold2. Train energy and dipole models with active learning3. Run equilibrium NPT molecular dynamics4. Run non-equilibrium transition trajectories4. Train the two-branch CNN6. Network inference and ranking statistic, including time-shifted background4. Train six ANN models5. Search hidden-layer size and transfer function6. Predict USECF with trained models3. Train source CNN classifiers on ZTF data4. Transfer-learn target classifiers on ILMT data5. Evaluate models on independent test data6. Deploy classifiers on full-frame ILMT images3. Run six previously trained classifiers on the observed cube4. Classify photometric cluster members with CNN3. Train the three regression architectures4. Predict DM for held-out synthetic spectra6. Apply trained models to real CHIME/FRB bursts2. Prompt LMs for ranked precursor sets3. Prompt LMs for calcination and sintering temperatures5. Generate synthetic solid-state recipe dataset6. Train SyntMTE and baselines in two stages7. Extract LLZO reference conditions from literature corpus8. Predict sintering temperatures for doped LLZO3. Train 1D CNN inversion model with leave-one-out folds and weight averaging4. Infer global oxide abundance maps5. Ablation and baseline model comparison under LOOCV5. Train charge-prediction neural networks6. Predict atomic charges8. Measure scaling on long alkane chains2. Train VIBANN surrogate3. Sample candidate latents and decode compositions4. Refine latent seeds by gradient optimisation7. Atomistic simulation of designed glasses8. Attribute hardness to elements and latent directions4. Fit MSORF regression coefficients5. Optimize hyperparameters on leave-one-out errors6. Predict properties of test molecules3. Identify and mask outlier intensities4. Cluster lines into homogeneous subsets5. Size architecture with PCA and augment inputs6. Fit surrogate models to the training set7. Predict line intensities at test points1. Predict full-length honey bee Vg structure5. Train Gaussian process regression charge model6. Predict charges across large RMC configurations1. Obtain AlphaFold models of human BRICHOS proteins5. Predict amyloid propensity of mature polypeptides6. Score pathogenicity of the T171I variant4. Classify single-cell fluorescence phenotypes7. Compute sequence- and structure-derived protein features8. Test features for discrimination between aggregation classes2. Encode residues with ESM-2 embeddings3. Train local–global dual-attention Transformer classifier4. Train DQN reinforcement-learning classifier6. Predict binding probability per residue on test sets3. Train FFiTrNet and baseline models4. Predict densities for held-out test split5. Predict densities for Huang & Massa compounds2. Obtain AlphaFold structural models with pLDDT and PAE8. Add sequence-feature predictions to protein web pages4. Train PepSeP1 and PepSeP6 networks5. Design peptide sequences for held-out and perturbed backbones3. Encode sequences4. Train per-residue classifiers5. Predict RCL positions on held-out sequences2. Predict pair structures with AlphaFold-Multimer4. Build structural and omics feature vectors for each pair5. Train random forest classifiers with feature pruning6. Score genome maintenance matrix and proteome-wide screens with SPOC4. Train 3D-CNN binary classifier5. Evaluate classifier on held-out test data6. Visualise model attention with Grad-CAM7. Score Schiff base structures from the CSD2. Predict segment structures with AlphaFold27. Predict nsP assemblies and inter-domain contacts with AlphaFold28. Predict SARS-CoV-2 replication complex as a control2. Pre-relax structures with a machine-learned potential5. Train invariant and equivariant GCNN energy models6. Predict Ehull across composition and ordering sets7. Evaluate pre-trained universal interatomic potentials4. Train one model per topology on both properties5. Predict adsorption properties on held-out structures6. Attribute adsorption to individual pores7. Inverse design with genetic algorithm2. Predict full-length models as reference4. Predict structure of each fragment6. Run sequence-based coiled-coil predictors5. Train multi-output surrogate neural network6. Solve MILP with embedded network for Pareto-optimal sequences3. Predict complex structures with five models6. Re-run models under input and attention ablations5. Optimise BSIM parameters with DDPG agent2. Encode domain sequences as language-model embeddings3. Train CATHe classifier and baseline models4. Evaluate on held-out test splits6. Predict CATH superfamilies for Pfam domains3. Label frontal waves and build augmented training set4. Train YOLOv3 frontal wave detector5. Evaluate detector on held-out and full-year imagery6. Apply detector to eleven years of imagery2. Predict per-chain structures with AlphaFold34. Train multi-task encoder-decoder with feature pyramid network5. Predict backbone atoms, Cα atoms and amino acid types3. Co-predict LC8–client complex structures with AlphaFold5. Train composite score threshold classifier5. Fit regression models to process-property data6. Predict tensile strength, modulus and hardness4. Train CrysFormer on map triples5. Predict completed electron-density maps3. Predict per-residue biophysical features from sequence4. Train and test random forest residue classifiers5. Classify MFIB residues with the combined model6. Derive surrogate rule models for interpretation5. Train query-aware ridge models and predict SuperCon Tc7. Predict Tc across Materials Project2. Predict structure and score solenoid class4. Redesign sequences and test designability6. Test in silico metrics against experimental outcomes7. Design terminal caps for beta-solenoids and filter2. Fit moment tensor potential3. Enrich training set by active learning4. Train neuroevolution potential5. Evaluate energy and force errors on held-out set6. Run machine-learning molecular dynamics2. Predict structures of human and yeast homologs4. Classify substitutions as driver or passenger1. Train one ML force field per XC functional2. NpT equilibration to obtain liquid densities3. NVT production molecular dynamics2. Cluster raw functional annotations into categories with an LLM4. Encode sequence, structure graph and function text into features5. Train M3Site and retrained baseline models6. Predict residue-level active site classes on held-out and unseen proteins4. Factorize the XRD map into representative bases5. Label phases probabilistically with CrystalShift6. Fit the Gaussian process phase map7. Choose the next anneal by maximizing logEI4. Score functional similarity of pairs with the SNN6. Check sibling labelling against experimental EC annotations7. Run annotation tools and embedding models over the orphans3. Train three conditional diffusion models4. Sample synthetic solar images per flare class5. Score generated images with cluster metrics, FID and classifier F16. Augment classifier training set with synthetic images7. Train binary 24-hour flare predictor with DDPM augmentation3. Train and tune seven learners per polymer4. Predict diameter for user settings5. Attribute parameter influence2. Re-predict REPI structure with AlphaFold21. Build initial reference dataset2. Train MLIP committees and optimise hyperparameters3. Run reactive scattering molecular dynamics4. Flag high-uncertainty structures during dynamics5. Cluster and sparsify selected structures3. Fit Extra-Trees surrogate on current dataset4. Score composition grid by expected improvement5. Cluster candidates and shortlist for synthesis7. Attribute predictions to elemental descriptors3. Characterise GMM behaviour on simulated member and field populations4. Fit two-component GMM and assign membership probabilities3. Meta-train emulator with first-order MAML4. Train comparison emulators5. Fine-tune emulators to novel redshift distribution6. Test emulated spectra against theory7. Run MCMC inference with emulators in the likelihood4. Stage 1: train five baskets of regression models6. Stage 2: refine with boosting and ensembles7. Predict emission behaviour for 1.25–2.0 wt% graphene3. Train generative galaxy image model4. Generate realistic galaxy images5. Train CNN morphology estimator6. Fit structural parameters with five codes3. Fit cluster expansion and sample toward ground states4. Train neural network potential with committee-driven dataset growth5. Rank enumerated configurations at unseen compositions6. Compare potential and cluster expansion against DFT7. Simulated annealing search for low-energy Cr(1-x)S structures8. Vacancy diffusion Monte Carlo on a strained CrS2 slab3. Train GAP interatomic potentials4. Iteratively generate and label new configurations6. Run production biased molecular dynamics7. Integrate gradients to free energy surface and rate constants8. Test TST assumptions with unbiased trajectories5. Self-supervised denoising pre-training of the CG graph encoder6. Train from scratch and fine-tune task models7. Predict complex binding affinity and interface type2. Embed wild-type proteins and retrieve similar datasets3. Generate MSA-based pseudo labels4. Meta-train PLM on auxiliary tasks with LoRA5. Fine-tune on target data with listwise ranking loss6. Score held-out mutants and benchmark against baselines7. Select Phi29 single-site mutants for testing3. Extract ProtT5 embeddings for lysine residues4. Train base modules and alternative ML/DL models5. Select architectures and hyperparameters by cross-validated grid search6. Train stacked meta-classifier (LMSuccSite)7. Score independent test set and compare with existing predictors8. Inspect learned features and training-size sensitivity3. Train predictive neural network4. Predict responses for free-form meta-atom pool4. Train XGBoost ensemble on degron features (MetaDegron-X)5. Train hybrid sequence-only deep network (MetaDegron-D)6. Evaluate by cross-validation and independent testing7. Compare with Degpred on β-TrCP2 substrate dataset8. Serve batch degron prediction and annotation via web server3. Encode sequences with ProtBert4. Fuse features and fine-tune the classifier5. Predict Kla sites on the independent external set6. Compare against ablations and published predictors7. Interpret model via permutation importance and attention maps3. Fine-tune ProtBERT with a regression head per property4. Predict properties for held-out sequences6. Inspect learned representations and re-split by cluster3. Train per-atom GPR models with adaptive sampling4. Fit Lennard-Jones nonbonded parameters against α-phase properties5. Optimise α and β crystal structures with FFLUX6. Lattice dynamics, free energies and IR spectra from FFLUX forces4. Train classifiers and tune hyperparameters5. Classify held-out test sequences6. External validation on an independent dataset2. Predict known complexes with three methods4. Generate de novo peptides for three targets6. Redesign sequences for selected backbones7. Repredict designed peptides3. Encode multi-view token sequences4. Build context visual graph for attention guidance5. Train encoder and decoders with view-masked self-supervision6. Segment volumes and pick particles2. Train ANN spectral interpolator5. Fit observed spectra to derive labels6. Internal accuracy tests on held-out synthetic spectra8. Estimate label uncertainties by repeated retraining and refitting3. ML-based map modification4. Train the moment tensor potential5. Validate potential against DFT7. Run tensile CMD on monolayer and nanotubes8. Run heating CMD to find phase transition3. Train property-conditioned diffusion score function4. Generate structures by conditioned reverse-diffusion denoising5. Refine denoising trajectory with Hamiltonian Monte Carlo on the learned noise energy3. Encode lightcurves as wavelength-time heatmaps4. Train classification and redshift CNNs5. Predict SN types and redshift PDFs6. Match supernovae to host galaxies and estimate host photo-z8. Pretrain with generic augmentations and fine-tune with targeted augmentations3. Fine-tune pre-trained CNNs on encoded images4. Classify held-out encoded sequences6. Quantify prediction uncertainty by Monte Carlo Dropout3. Fit Gaussian process surrogate on top of MLP prior4. Relax random structures on the surrogate surface7. Baseline random structure search with the MLP alone2. Train pixel-only classifiers3. Fit approximate PSF models5. Train PSF-aware SVM classifier6. Assign spectral class and SED template to stars7. Retrain PSF model with classified stars3. Predict per-residue intrinsic disorder5. Predict liquid–liquid phase separation propensity1. Train and test ALINet on MNIST as an architecture check4. Train ALINet on RIAF images5. Train InvNet to map parameters to latent distributions6. Recover physical parameters from held-out images7. Generate images directly from physical parameters3. Train Random Forest classifiers5. Classify unclassified white dwarfs into DA and non-DA6. Classify non-DAs into primary and secondary subtypes4. Select features with gradient boosting5. Train multichannel CNN classifier6. Predict submitochondrial localization8. Cluster simulation trajectory residues and compute mutual information3. Encode local atomic environments as descriptors4. Train the interatomic potential5. Run void-nucleated melting simulations6. Run two-phase coexistence simulations8. Compute caloric and thermal expansion curves and compare with reference data5. Train the two CNN models6. Predict background-fit and confidence-line parameters for every pixel3. Self-supervised pretraining of the HSMP encoder4. Fine-tune and benchmark on nine property tasks6. Predict Tg across the library with a ten-model ensemble4. Train GNN classification and segmentation models5. Predict overall DNA/RNA binding function6. Predict residue-level NA binding sites7. Attribute predictions to features and surface regions3. Predict structures for sequences lacking database structures4. Retrain sequence classification models on Task 1 train split5. Train contrastive reaction-enzyme retrieval models6. Classify held-out sequences by EC number7. Retrieve EC numbers for held-out query reactions3. Fit kernel ridge regression models4. Predict quantum properties3. Train interatomic potential as denoising model5. Relax on pseudo energy surface (denoising)4. Generate ESM-2 protein embeddings5. Train and tune classifiers on each feature space6. Apply locked models to test, control and external sets7. Benchmark against existing AMR detection tools8. Structural and phylogenetic error analysis of persistent misclassifications3. Train a single VAE on all compositions4. Sample latent space and decode candidates7. Relax structures and predict energy with ML potentials4. Encode antibody sequences as ESM-2 residue embeddings5. Train chimeric classifier and quantitative regression head6. Predict hot spots and extents on independent antibody set8. Screen clone panel from sequence alone4. Fit E_HOMO–E_ox correlation and predict E_ox5. Derive molecular descriptors and predict solubility6. Train and evaluate ML models on OxPot7. Attribute feature importance in the trained GIN4. Select feature subset with SVM-RFE + CBR5. Train two-layer stacking classifier6. Evaluate by cross-validation and independent test3. Aggregate volunteer votes and markings into labels4. Train AutoML object detection model5. Run detector over the archive3. Train domain-adapted CNN ensembles for global morphology4. Classify galaxies into four global morphological classes5. Fine-tune Zoobot foundation model on Galaxy Zoo CEERS labels6. Predict featured, edge-on and bar probabilities3. Train artificial neural network on the design matrix4. Predict CNP yield with the network and compare errors against the response surface model4. Train random forest and neural network ZT regressors per temperature5. Predict ZT for held-out test points7. Attribute ZT predictions to transport features with SHAP3. Fine-tune off-the-shelf architectures and retrain baseline4. Score test-split alerts with trained models6. Embed latent representations with UMAP7. Flag outliers with Isolation Forest3. Obtain stellar number density field from Gaussian Process regression4. Supply missing radial velocities with Bayesian neural network predictions2. Train SNR interpolator network3. Predict SNRs and estimate selection function by Monte Carlo4. Train selection function network6. Sample hyperposterior with selection correction2. Generate TCR sequence designs with ProteinMPNN and ESM-IF15. Model designed sequences with TCRModel24. Train fully-connected networks to predict peak halo mass5. Search architectures and hyperparameters on validation loss6. Predict halo masses for the held-out simulation box8. Retrain with inputs masked to locate the informative features2. Benchmark NNP host-gas interaction energy against DFT and DFTB4. Run NNP molecular dynamics sampling4. Train the modular set-based network and its conditional-VAE module5. Search hyperparameters and select a model from the Pareto frontier6. Predict the electron pressure field for held-out clusters8. Ablate and add modules to interpret what the network uses4. Predict the structure of full-length NorD5. Predict NorQ-NorD complex models2. Train machine-learned force field3. Committee-based active-learning selection and DFT labelling5. Run MLFF molecular dynamics across temperatures3. Train regressors mapping observables to disk parameters4. Infer disk parameters for archival ALMA sample3. Pre-train equivariant network on masked residue identity5. Zero-shot scoring of mutational effects6. Amortize relaxation by fine-tuning on relaxed predictions7. Fine-tune end-to-end on measured stability and binding effects8. Rank substitutions at viral antigen sites4. Direct-learning training on RPA spectra5. Transfer-learning retraining from the IPA model6. Predict spectra for held-out materials8. Predict IPA-RPA similarity and map latent space3. Train normalizing flow on stellar phase space4. Sample flow and auto-differentiate distribution-function gradients5. Train potential network and frame rotation speed6. Evaluate accelerations and total density from the potential3. Train pairwise binary CNN classifiers4. Hyperparameter search and threshold selection6. Train single multi-class CNN classifier7. Classify bursts and combine binary confidences8. Explain predictions with SHAP3. Model miscentering and compute tangential shear profiles4. Train set-based summary network5. Compress catalogs into neural summary vectors6. Train masked autoregressive flow posterior estimator7. Infer posteriors for mock catalogs4. Train graph-encoder/transformer-decoder variational model per system5. Sample latent space to generate and interpolate backbones3. Train U-Net models to map mass to observables4. Predict observable maps for held-out test clusters5. Predict observable maps from dark-matter-only simulations7. Embed final convolutional layer for interpretation2. Extract protein language model embeddings and physicochemical features3. Select optimal feature groups by incremental feature selection4. Select base and meta learners5. Tune learner hyperparameters by grid search6. Train stacking ensemble on imbalance-corrected data7. Predict glycosylation sites on independent sequons8. Attribute feature-group importance with SHAP4. Train and compare classifiers6. Classify survey objects with BROWDIE7. Estimate effective temperature by regression4. Train one U-Net per loss function and pick hyperparameters5. Segment held-out test maps2. Train MACE potential with active learning3. Run production MD for five slit widths and both defects4. Run temperature series for the 3L system6. Check free energy profiles with umbrella sampling3. Train CNN lens/non-lens classifier in two phases4. Train CNN regression network for source redshift5. Score blind quasar samples with the classifier6. Estimate and refine background source redshift4. Train DeepClean per science segment5. Offline noise prediction and subtraction6. Low-latency cleaning of 1 s frames2. Train kinematic base network NN_FIRE5. Train parallel dynamics-based network NN_parallel6. Classify the Gaia DR3 target sample2. Predict per-residue intrinsic disorder3. Predict disorder-based binding regions (MoRFs)4. Predict phase-separation propensity4. Train MLP and CNN classifiers6. Classify held-out test orbits8. Apply classifier to full MAVEN record2. Extract ESM2 attention matrices and residue embeddings6. Train Word2Function on word embeddings to predict GO terms7. Attribute words to GO terms and populate WordTableGO658. Evaluate coverage and GO prediction against baselines3. Select features by random forest importance4. Train and tune regression models5. Evaluate predictions against held-out DFT data6. Compare predictions with reported experimental systems2. Screen descriptors with correlation analysis and SHAP3. Pretrain BPNN and run two continual-learning stages4. Evaluate forward model against baselines and literature data6. Score candidates with the surrogate model2. Select ANN architecture and train on bootstrap resamples3. Reconstruct H(z) as ANN ensemble average4. Leave-one-out cross-validation of the reconstruction5. Infer BNN weight posterior with NUTS6. Reconstruct H(z) by marginalising the BNN posterior3. Reduce features and normalise4. GA hyperparameter search and base-model selection5. Generate virtual sample inputs with GMM6. Generate virtual sample outputs with RegGAN7. Retrain base model on augmented training set8. Evaluate on test set and compare VSG methods2. Predict band gaps from structure with AFLOW-ML5. Reduce dimensionality and train four classifiers per approach6. Classify unlabelled materials as suitable or unsuitable3. Encode enzymes and substrates4. Train the bilinear attention and evidential model5. Predict kcat/Km with uncertainty6. Evaluate against held-out data and a baseline7. Interpret attention over residues and substrate atoms3. Score mutations with potential‐like methods4. Fit mass‐balance correction coefficients5. Apply corrected scores to test sets and to other methods2. Train the PFP potential3. Compute lithium diffusion barriers in LiFeSO4F4. Optimise MOF cells and compute water binding energies5. Monte Carlo simulation of Cu–Au ordering6. Compute methanation reaction barriers on Co(0001)7. Screen promoter elements for CO dissociation2. Detect radio sources and infrared hosts with Gal-DINO4. Adapter fine-tune OpenCLIP5. Embed extended EMU sources into searchable database6. Encode user text or image query3. Train neural network emulators4. Assess emulator accuracy against CAMB6. Run MCMC inference with the emulators3. Fit PCK surrogate4. Evaluate surrogate on held-out test set5. Surrogate-based MCMC posterior sampling7. Coverage study on synthetic planets4. Train joint-property model with linear mixing operator5. Predict properties for held-out species and blends7. Search latent space for mixtures matching targets4. Generate candidate reduced schemes with genetic operators5. Score candidates against full model and evolve next generation4. Encode features and interface graph5. Train Siamese graph convolutional network6. Tune hyperparameters under protein-level cross-validation7. Predict mutation effects on held-out and multiple-mutation sets8. Evaluate against measurements, baselines and ablations4. Train CNN classifiers on real and synthetic images5. Score SDSS galaxy catalog for polar ring pattern4. Fit scale network and structure factor posteriors by variational inference5. Infer merged structure factor amplitudes3. Predict complexes with AlphaFold v2.2 and v2.34. Predict complexes with ColabFold and output each recycling iteration5. Model unbound antibody and antigen subunits7. Re-run AlphaFold with custom templates and without MSAs3. Tune and fit seven multi-output regression models4. Predict colour outputs and score models6. Bayesian optimisation of bleaching conditions4. Train and tune classifiers6. Classify query helices with the ensemble3. Tune hyperparameters by grid search with cross-validation4. Train classifiers on general-relativistic waveforms5. Classify held-out waveforms and score over repeated splits6. Test GREP-trained models on general-relativistic waveforms7. Reduce dimensionality and inspect decision boundaries8. Sweep preprocessing choices and re-score accuracy2. Interpolate multi-band light curves with Gaussian Processes5. Build synthetic spectral templates with two-dimensional GP regression4. Train component neural networks5. Predict multipoles for held-out cosmologies7. Run mock full shape MCMC analyses2. Fine-tune MACE interatomic potential3. Train GNN band-gap model in two stages4. Sample chemical disorder, relax and rank configurations5. Compute phonons and run finite-temperature molecular dynamics6. Predict band gaps of MD snapshots and ensemble-average8. Mode-resolved electron-phonon analysis with frozen phonon distortions3. Train latent SDE on simulated light curves4. Reconstruct light curves and infer parameter posteriors5. Fit multitask GPR baseline to the same light curves4. Train SALTED electron-density model5. Test predicted densities, fields and forces against DFT6. Run data-driven QM/MM molecular dynamics4. Train base U-Net on full dark matter density field5. Curriculum-train on sparser subhalo density fields6. Predict structural segmentation over the volume3. Fit PCA baseline on training split4. Train convolutional autoencoder with IOB layer5. Encode and reconstruct held-out bursts with IOB-CAE6. Project CHIME-only data with PCA to find outliers6. Train CNN classifiers by transfer learning7. Classify held-out objects and compare models2. Embed perovskite composition in four dimensions3. Fit joint density and forward predictor4. Predict PCE from the conditional density5. Read cluster prototypes from mixture components6. Sample synthetic device records7. Infer synthesis conditions for a target PCE8. Refit the density on data with missing entries4. Train convolutional auto-encoder on unlabelled profiles5. Encode profiles into embedding features6. Train and score random forest explosion classifiers7. Propagate labels into reduced training splits and retrain5. Train neural network emulators of PHOEBE6. Fit observed photometry by nested sampling with the emulator5. Train sequence classifiers on simulated patterns6. Classify symmetry on simulated test splits7. Test trained models on experimental RRUFF patterns8. Peak-masking feature importance analysis2. Extract materials, properties, values and descriptors by prompting3. Classify each extraction with follow-up prompts6. Train graph neural network band gap predictors on each dataset7. Compare models by cross validation, shared hold-out set and matbench leaderboard8. Prompt the LLM to write and run code that trains predictors3. Fit mixture model to calibration velocities4. Score thick disk probability for Kepler stars5. Train EINN by backpropagation and convert weights to coefficients6. Refine coefficients with numerical Bayesian inference4. Train the CNN with a reweighted loss5. Reconstruct convergence maps for test, null and masked inputs8. Apply the CNN to Subaru/HSC Coma observations3. Train ProteinMPNN with backbone noise4. Generate sequences for fixed backbones with tied positions6. Predict structures of designed sequences with AlphaFold1. Neural network search for strong lens candidates2. Reject non-astrophysical alerts with braai4. Score candidates with BTSbot4. Pick filament segments with a neural-network picker3. Train deep neural-network potential4. Compute phase diagram with multithermal–multibaric simulations5. Simulate homogeneous nucleation of α-Ga and β-Ga6. Seeding runs to find critical nucleus sizes3. Label sites by binding-type and cluster GCN values into groups4. Apply frequency and coverage scaling factors to low-coverage data6. Train neural network ensembles for binding-type and GCN pdfs7. Infer microstructure from experimental spectra2. Rank precursor sets by DFT reaction energy4. Identify phases and weight fractions from XRD patterns7. Benchmark against black-box optimisers on the YBCO dataset4. Train and compare five base classifiers and a tuned RFC5. Resample dataset with SMOTE-Tomek links and retrain RFC6. Predict phases of unseen literature alloys with all five models7. Predict the phase of a new alloy composition with RFC4. Train activity classifier and titer regression model5. Predict untested chimeras and pick next batch by UCB7. Model block contributions to activity3. Evaluate posterior probabilities of orientation and class assignment4. Update 3D map, noise and signal power spectra7. Estimate orientational assignment accuracy from the model3. Train Bayesian CNN on spectra with APOGEE labels4. Evaluate on held-out test set5. Predict parameters for LAMOST DR8 giants3. Optimise architecture hyper-parameters4. Train ExoMiner and retrain baselines5. Score held-out TCEs and unlabelled KOIs7. Branch-occlusion explainability analysis8. Transfer Kepler-trained model to TESS signals5. Train CNN glitch classifier6. Classify glitches and assign confidence3. Train initial Random Forest classifier4. Score nightly alerts for early SN Ia probability7. Add new labels to training set and retrain2. Re-score alerts with real/bogus classifier4. Optimise hyperparameters and train BTSbot and uni-modal baselines5. Score alert packets with BTSbot7. Obtain and automatically classify spectra3. Train regressors and select regressor:selector pair4. Predict hysteresis and uncertainty across search space4. Self-supervised pre-training on 3D structures5. Fine-tune with gas and operating-condition blocks6. Predict uptake and structural features for unseen materials4. Train Mask R-CNN on manually annotated slices5. Segment individual particles slice by slice2. Fit moment tensor potentials to current database3. Sample new environments by uncertainty-driven MD and amorphous matrix embedding5. Fit final non-linear ACE potential6. Run large-scale ACE molecular dynamics for silica, surfaces, aerogels and SiO4. Segment measured materials into phase regions5. Propagate phase labels and property estimates to unmeasured materials7. Select the next material to measure and the optimum to pursue2. Propose precursor sets and synthesis temperature5. Identify phases in diffraction patterns6. Confirm phases by automated Rietveld refinement7. Propose improved reaction routes with ARROWS4. Train GP classification and regression models5. Predict across the library and select exploration and verification sets6. Select optimal chimeras and CsCbChR1 block swaps by confidence bounds8. Weight sequence and contact features for interpretation2. Train dataset-specific crYOLO network3. Pick particles across full datasets5. Train general network on pooled datasets6. Pick previously unseen datasets with general network2. Train AlphaFold on PDB structures3. Predict structures for unlabelled Uniclust sequences4. Retrain with self-distillation and fine-tune5. Predict structures and estimate confidence8. Probe network behaviour with per-block structure modules3. Fit maximum entropy model and infer residue couplings8. Run alternative coupling methods through the same pipeline3. Pick and trace thick and thin filaments5. Predict atomic structures of filament components7. Denoise representative tomograms1. Benchmark model and acquisition functions on P450 data3. Fit agent's landscape model to observed data4. Predict activity and thermostability across the space8. Unified landscape model for analysing agent behaviour3. Train neural-network particle picker on manual picks4. Auto-pick particles with the trained network6. Predict subunit complexes with AlphaFold2 Multimer7. Build, refine and assemble atomic models3. Train peak picking, deconvolution, shift prediction, ranking and density models4. Visual spectrum analysis: pick and deconvolve cross-peaks5. Unalias folded peaks to true resonance frequencies6. Assign chemical shifts and refine with GNN predictions7. Calculate structure proposals and select with GBT2. Model domain structures and trim flexible termini3. Design de novo and redesigned sequences5. Infer K50 and folding stability, then filter for quality6. Principal component analysis of per-site amino acid stabilities7. Fit classifier predicting wild-type amino acid from stabilities8. Identify functional sites using evolutionary sensitivity scores
209records
70structural-biology
70materials-chemistry
69astronomy

In 166 of 179 studies the AI produced the result; in the rest it played a supporting part. 88 of 179 papers say their code is available. Both counts are over the papers drawn at random, not the ones chosen for the collection. How we decide what to include.

git log --latest

aix-00279

Machine learning predicts where dopant atoms sit in hexaferrite crystals before refinement

classificationproperty prediction

materials-chemistry · ai-result · Advanced Science · 2026

aix-00278

Neural networks sort millions of stars to build a young-star candidate catalogue

classification

astronomy · ai-result · Astronomy and Astrophysics · 2026

aix-00277

Software reads AlphaFold predictions to infer how protein complexes assemble

structure determination

structural-biology · ai-result · Frontiers in Bioinformatics · 2026

aix-00276

Machine learning sorts a simulated polymer into ordered and disordered beads

classification

materials-chemistry · ai-result · arXiv · 2023

aix-00275

AI model trained on protein folding predicts how mutations affect binding

property prediction

structural-biology · ai-result · iScience · 2026

aix-00274

Neural network trained on simulated data hunts solar wave trains in radio spectra

detectionclassification

astronomy · ai-result · Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences · 2026

aix-00273

Deep learning model flags bacterial proteins likely to drive infection

classificationstructure determination

structural-biology · ai-result · Bioinformatics · 2026

aix-00272

Tree-based models predict sliding wear of carburised and plain engineering steels

property prediction

materials-chemistry · ai-result · Materials · 2026

aix-00271

Machine learning predicts galaxy 'relicness' from survey measurements, cutting spectroscopy demand

property predictionclassification

astronomy · ai-result · Monthly Notices of the Royal Astronomical Society · 2026

aix-00217

Reconstructing the early universe's density field with a smoothed map of cosmic structure

structure determination

astronomy · ai-result · Astronomy and Astrophysics · 2026

all articles → · browse by field and method →