A useful calculation starts with an observable, a physical model and a way to test it. These methods make their assumptions, numerical checks and research foundations inspectable.
One-dimensional steady conduction with specified layer conductivities and interface conductances.
Analytical Fickian diffusion into a homogeneous semi-infinite solid.
Ideal configurational entropy with a constant symmetric interaction energy.
Small closed-shell molecules containing H through Ne with RHF or PBE and a declared basis.
Monatomic FCC, BCC or simple-cubic crystals using explicit pair-potential parameters.
A bounded Gaussian-process delta model for one system, observable and explicit low/reference method pair.
Neutral closed-shell cells containing H through Ar with fixed PBE and GTH-PBE pseudopotentials.
Monatomic neutral defects under a specified Lennard–Jones model, with immutable before/after structures.
Does the implementation satisfy its equations, invariances and convergence checks? An analytically known limit can answer this without a material experiment.
Does it reproduce an independently specified calculation under the same conditions? Basis, functional, potential and normalization must be compatible.
Does the model predict independent measurements under matching conditions? A numerically converged result can still use an inadequate physical model.
Does the result distinguish competing explanations or change the next experiment? A low average error does not answer every research question.
24 research papers connect scientific agents, electronic structure, learned potentials, uncertainty and validation. The notes below describe selected sections inspected; they are not independent reproductions or endorsements of every claim.
Active learning
Treat uncertainty, diversity, exploration and retraining as separate choices. An uncertainty threshold alone does not establish accuracy.
Read in context: Sections 2.1–2.2: scheduling and acquisition.
Read the paper ↗Scientific harness
Represent scientific assets and dependencies explicitly; a plan revision must identify which prior artifacts remain valid.
Read in context: Architecture and stateful workflow sections.
Read the paper ↗Scientific harness
A useful materials agent needs structure handling, computation, convergence and interpretation in addition to retrieval.
Read in context: Section 2.1 architecture and geometry workflow.
Read the paper ↗Scientific harness
A physics skeptic should test conserved quantities and scaling laws appropriate to the modeled system.
Read in context: Sections 2.2–2.8 plan, execution and skeptic design.
Read the paper ↗Numerical verification
Exact invariances test internal consistency without pretending to measure experimental accuracy.
Read in context: Certification overview and symmetry/integrability families.
Read the paper ↗Lattice dynamics
Finite supercells limit wave-vector resolution; primitive-cell response offers a potential acceleration requiring a differentiable periodic representation.
Read in context: Section II.1 force response and dynamical matrix.
Read the paper ↗Long-range physics
Short-range local energies need explicit long-range treatment for field response; slab and bulk electrostatics differ.
Read in context: Section II.1 charge projection and electrostatic decomposition.
Read the paper ↗Training data
Training coverage must include off-equilibrium and low-coordination environments; element count alone overstates transferability.
Read in context: Dataset construction, dimers and low-coordination coverage.
Read the paper ↗Scientific harness
Chemically valid analogies require matching electronic context. Converged wavefunctions still need physical-state checks.
Read in context: Sections II.2–II.4 precedent, audit and convergence.
Read the paper ↗Scientific harness
General composable scientific actions and task-specific literature are more useful than a narrow task script.
Read in context: Architecture, procedural memory and guided research.
Read the paper ↗Evaluation
Partial scores and confident prose do not establish completion; test the whole requested artifact bundle.
Read in context: Task contract and complete-bundle evaluation.
Read the paper ↗Active learning
Diverse training configurations can improve energy errors while worsening force/stress errors on a different distribution.
Read in context: Reported held-out behavior and acquisition motivation.
Read the paper ↗Active learning
Select data for a target observable rather than indiscriminately reducing every parameter uncertainty.
Read in context: Section 2 and quantity-of-interest motivation.
Read the paper ↗Lattice dynamics
Phonons depend on local curvature and may require targeted reference data even when energies appear accurate.
Read in context: Fine-tuning, force accuracy and property propagation.
Read the paper ↗Foundation potentials
Conservative force behavior and task-dependent speed/accuracy tradeoffs matter alongside parameter count.
Read in context: Approach, model family and conservative refinement.
Read the paper ↗Training data
Diverse sampling and consistent electronic-structure settings change the quality of a learned potential.
Read in context: Dataset construction and electronic-structure consistency.
Read the paper ↗Thermal transport
Random interpolation can hide failure on new symmetries or low-conductivity targets.
Read in context: Benchmark design and random/space-group/property splits.
Read the paper ↗Domain adaptation
Dynamics-sensitive applications require relevant configurations and independent experimental evidence; more augmented data can degrade downstream performance.
Read in context: Training composition and benchmark extraction sections.
Read the paper ↗Domain applicability
Bulk performance does not establish quality for molecules, surfaces or mixed-dimensional interfaces.
Read in context: Benchmark setup and conservative/direct-force model table.
Read the paper ↗Active learning
The labeling budget and stopping condition are part of the scientific method, not operational afterthoughts.
Read in context: Pool-based algorithm and application challenges.
Read the paper ↗Sampling
Diversity alone does not guarantee physically representative samples; biased dynamics changes the sampled distribution.
Read in context: Enhanced-sampling motivation and distribution constraints.
Read the paper ↗Scientific memory
A useful scientific memory distinguishes run observations, project lessons and reviewed reusable protocols.
Read in context: Three-tier memory and promotion policy.
Read the paper ↗Evaluation caution
The inspected feature list includes energy while the target is formation energy; the text also mixes paired/unpaired test descriptions. These require clarification before adopting reported gains.
Read in context: Sections 3.1–3.5 features, splits and statistics.
Read the paper ↗Mechanical properties
Elastic response depends on second derivatives; energy and force accuracy cannot substitute for an elastic-property benchmark.
Read in context: Motivation and methodology.
Read the paper ↗