Bitcoin 1.01 — Science Layer (draft)

A design that directs half of the computing power spent on mining to drug-candidate screening and materials discovery. Status: DRAFT — not in code. Prerequisites: launch, Phase 6 (encrypted P2P), the archive layer (TASARIM.md section 11), addresses with small signatures (ML-DSA).

Sources were surveyed on 2026-10-26. Figures and licences marked (to be verified) were seen only in secondary sources; check them against the official page before use.

1. Core principle: "half of the computing power" = half of the reward

The chain cannot see or prove how a computer divides its processor. That is why the split is made economically:

• Consensus rule: while a science campaign is active, 50% of every block's reward goes to the "science pool" and 50% to the miner. The pool pays those who do verified science work.
• Local setting: a "Mining ↔ Science" slider in the app (default 50%). This is not a rule but the person's own thread / graphics-card split on their own computer. People shift their power toward whichever side pays better; the two reward pools settle into rough balance.
• The honest cost: when the mining reward halves, the cost of taking over the chain also roughly halves (ASERT adjusts difficulty to the real power). That is why the ratio is kept as a single constant (BILIM_PAYI); it can start lower while the network is small and be raised to 50% later.

Why mining itself cannot do scientific computation (in short): the block puzzle must be instantly verifiable, must derive from the block (no one may choose it), every attempt must be an independent chance, and it must have no value outside the chain. Scientific computations do not meet these four conditions. The known secure example in academia (Ofelimos, CRYPTO 2022) works only for special combinatorial optimization problems. Hence two layers: proof of work for security, and a separate, verifiable science layer for usefulness.

2. Suitable work

Scale examples (rough):

• D1, 1 billion molecules, CPU only: tens of core-seconds per molecule → hundreds of core-years. 1,000 computers giving half their power (4 cores, 2 for science) finish in about 3–4 months. With D4, screening 1–5% of the library takes a few days.
• D1, graphics cards: the Uni-Dock paper screened 38.2 million molecules in 12 hours with 100 V100 GPUs; that is ~9 molecules per second per GPU. 1 billion molecules ≈ 1,300 GPU-days; ~1 month with 100 part-time gaming GPUs.
• M2, MOF screening for CO2: the 8,300 public structures of CoRE MOF 2024 are a bounded, finishable job; an ideal first campaign to test the infrastructure.

3. Datasets and sources

3a. Drugs: molecule libraries

3b. Drugs: targets and diseases

3c. Drugs: activity data (parameter choice, validation, model training)

3d. Drugs: software

3e. Materials: datasets

3f. Materials: models and software

3g. Volunteer computing and proof-of-useful-work literature

• BOINC: every job runs on at least two unrelated computers and is accepted if the results agree. For floating-point differences, a tolerant validator or "homogeneous redundancy" (sending to the same operating system and processor vendor) is used. This is the basis of our verification.
• Ofelimos (CRYPTO 2022) and FRLS (2025): provably secure proof of useful work; only for optimization problems solved by local search.

4. How the datasets are connected

4a. Campaign manifest

Every piece of science work starts with a campaign. The manifest is a JSON document signed with SLH-DSA by one of the science board's keys (field names are Turkish, as in the program):

{
  "kampanya": "amr-2027-01",
  "amac": "Carbapenem-resistant A. baumannii — NDM-1 metallo-beta-lactamase inhibitor",
  "tur": "D1",
  "hedef": { "kaynak": "PDB:xxxx", "hazir_dosya_sha256": "…", "kutu": [x, y, z, dx, dy, dz] },
  "kutuphane": { "kaynak": "ZINC-22 3D", "surum": "…", "dilim_boyu": 1000, "dilim_sayisi": 1000000,
                 "dizin_merkle_koku": "…" },
  "yazilim": { "ad": "vina", "surum": "1.2.x", "ikili_sha256": "…" },
  "parametreler": { "exhaustiveness": 8, "tohum": "ORION-512(kampanya || dilim || molekül)" },
  "dogrulama": { "k": 2, "skor_yuvarlama": 0.01, "poz_yuvarlama_A": 0.001 },
  "odul_birim": "0.5",
  "sonuc_lisansi": "CC0",
  "bitis_yuksekligi": 0
}

4b. Preparation (campaign team, once)

This step needs experts and is not distributed:

1. Target selection (WHO list, TDR Targets, a partner laboratory's suggestion).
2. Protein preparation: protonation, finding the binding pocket.
3. Redocking with known active molecules (ChEMBL, PDB) and an enrichment test with decoy molecules: if the parameters hold up, the campaign opens.

4c. Work units and data distribution

• Work unit = (campaign, slice number). Its content is deterministic: specific lines of the library file. The slice's correctness is proven against the Merkle root in the manifest.
• Openly licensed data (PDB, AlphaFold, ZINC if permitted): distributed between peers in content-addressed pieces (ORION-512) on the archive layer. Archivers are rewarded for keeping it.
• Data with restricted licences (such as Enamine): the network distributes only slice IDs and hashes. The participant downloads the data from its source, under terms they accept themselves.

4d. Computation and determinism

• The scientific software ships inside the Bitcoin 1.01 release, with pinned hashes. Campaigns cannot make the program download anything; the manifest only selects, by hash, a binary that came with the release.
• It runs in a separate, low-privilege process without network access (a sandbox).
• Single thread, fixed seed, canonical output: scores are rounded to 0.01 kcal/mol, coordinates to 0.001 Å, and the result file is hashed.

4e. Verification

1. Assignment: k people are chosen by a stake-weighted lottery from the hash of the block in which the slice opens and the slice number (the lottery of TASARIM.md section 11). No one can choose which slice they get.
2. Commit → reveal: the participant first writes the commitment ORION(result || address) to the chain and reveals the result when the window closes. Copying someone else's result does not work.
3. Comparison: deterministic jobs exactly, jobs like D3 / M3 with tolerance (BOINC style).
4. Cheap checks: any node can rescore the claimed best poses (D1) or compute a single point on the final structure (M1 / M2).
5. Penalty: a deduction from the stake for a disagreeing result; removal from the campaign if it repeats.

4f. Publishing and tracking the results

• Only a summary and a Merkle root stay on chain; the full results (SDF / CSV / pose files) live on the archive layer.
• A "Science" page in the explorer: campaign progress (percentage of slices finished), the best candidates (SMILES, score, downloadable pose), number of participants, verification rate, disagreements.
• Results are public under CC0. Bulk exports are handed to partner institutions at regular intervals.
• Laboratory loop: a partner laboratory orders the best candidates (Enamine make-on-demand synthesis) and tests them; active/inactive results are written to the archive and the chain and become training data for the next model.

5. Adding it to mining (consensus design)

Stages:

6. First campaign proposals

1. Infrastructure pilot, materials: CO2 / H2O adsorption screening (M2) on CoRE MOF 2024 (8,300 open structures). Data and tools are open; result verification is cheap; the job is bounded and finishable. Note: general-purpose ML potentials can be inaccurate for MOF adsorption; models trained on ODAC fit better, but check their licence.
2. Antibiotic resistance, drugs: a target from the WHO critical list (for example a resistance enzyme in carbapenem-resistant A. baumannii), D4 + D1 + D2. The candidates found can be synthesized and **tested for free at CO-ADD**; for a project without a laboratory, this is the shortest path.
3. Neglected disease: a target prioritized by TDR Targets, together with an institution such as DNDi.
4. Battery electrolytes: oxidation stability screening (M3) for molecules within the scope of LIBE / OMol25.

7. Security, ethics and law

• Software supply chain: only programs that come with the signed release, with pinned hashes, run on volunteers' computers; the sandbox is mandatory. A campaign must never be able to ship code.
• Dual use: a screening platform can also be pointed at a harmful purpose. Campaigns open only with therapeutic goals and science board approval; manifests are public and open to review; no campaign aimed at increasing toxicity is accepted.
• Licences: data derived from ChEMBL must be shared under CC BY-SA; CARD requires a licence for commercial use; nothing from Enamine / ODAC / OMol25 should be distributed without reading their terms. The results we produce are CC0.
• Expectations: experimental hit rates in docking screens are low; ML materials predictions must be validated by DFT and experiment. The platform produces candidates, not drugs or materials. Announcements must say this clearly.
• Partners: for target selection and laboratory testing, a university medicinal chemistry / computational chemistry group, CO-ADD (antimicrobial testing), DNDi (neglected diseases); for materials, a computational materials group.

8. Campaigns without partners

Campaigns that need no agreement with any institution and no data licence. What they share: **we generate the input ourselves** (molecules and crystals are enumerated from a set of rules), open-source programs do the computing, and the results are published under CC0. This does not "discover" a drug or a material; it produces **computational reference data** that feeds other people's models and laboratories. Most of today's drug and materials AI is trained on sets produced this way, without a laboratory (QM9, ANI, SPICE, OC20, OMol25).

Shared infrastructure

• Molecules from numbers: a work unit is a range such as "molecules 12,000,000–12,000,999". Every participant generates the molecules on their own computer from the rules; there is no dataset to distribute. Number → mixed-radix number → (scaffold, substitution choices) → canonical SMILES. Numbers that break a rule or produce a molecule already generated are skipped deterministically.
• Software: RDKit (BSD), xtb (LGPL-3.0), PySCF (Apache 2.0), Quantum ESPRESSO (GPL), AutoDock Vina (Apache 2.0). GPL/LGPL tools are not embedded in the program; they run as separate processes. Versions and settings are pinned in the campaign manifest; mixed levels of theory make the data worthless.
• Publishing: the Merkle root on chain, the files on the archive layer; also Zenodo (DOI) and Hugging Face. A free account is enough for these; no partnership is needed. A dataset paper (arXiv / ChemRxiv) to get noticed.
• Licence: data CC0. The project's own code has no licence (there is no LICENSE file in the repository); a licence such as MIT or Apache 2.0 should be added before publishing it as open source.

8a. ORION-Electrolyte (proposed first campaign)

The molecular space of battery electrolytes: solvents, salt anions and additives.

Why it is valuable: the main obstacle in electrolyte screening is the voltage range within which a molecule does not break down. The large set in this field (OMol25) has gated access, and the open one, LIBE, is small (~17 thousand species). A CC0 set at a single, consistent level and at million scale fills a real gap. Time per molecule at this level may be a few core-hours; it will be measured with a prototype.

8b. ORION-Molecule

A quantum reference for small drug-like molecules: energies and forces (ωB97M-V/def2-TZVPD) for conformers of molecules generated by our own enumeration (inspired by the GDB method; GDB-17 contains 166 billion molecules). The aim is to train the ML potentials used in drug discovery; similar to SPICE / ANI, but entirely CC0.

8c. ORION-Binding

Docking scores and poses (Vina) pairing open protein structures from the PDB (for example enzymes of WHO critical pathogens) with molecules we generate ourselves, CC0. Valuable as ML benchmark data; candidates could later be sent to CO-ADD if desired. Because docking scores are noisy, their lasting value is lower than that of quantum data. PDB terms of use are at wwPDB (to be verified).

8d. ORION-Crystal (second stage)

Element substitution in known crystal types → relaxation with an ML potential (MACE-MP-0) → DFT check with Quantum ESPRESSO → reference energies for stability (convex hull). New stable inorganic material candidates for battery cathodes and catalysts. Postponed, because DFT is heavy.

8e. Mathematics

Collatz verification, prime gaps. The option that best fits the project's √2 / 3x+1 identity and is the easiest to verify; its social value is symbolic.

8f. ORION-Nucleus: the effect of the forces on the nucleus

A study that models the effect of the strong, electromagnetic and weak forces on isotopes, decays and reactions using measured masses. Tool: araclar/cekirdek_modeli.py (no extra libraries needed). Data: the AME2020 atomic mass table (IAEA, free).

python araclar\cekirdek_modeli.py mass_1.mas20.txt           # all parts (~3 s)
python araclar\cekirdek_modeli.py mass_1.mas20.txt deneme    # only the electromagnetic scaling experiment

Formula (the Bethe–Weizsäcker semi-empirical mass formula, MeV):

B(Z,N) = a_v·A − a_s·A^(2/3) − a_c·Z(Z−1)/A^(1/3) − a_a·(N−Z)²/A + δ(A)
         volume   surface       electromagnetic       symmetry        pairing

The volume and surface terms are the "liquid drop" counterpart of the strong force; a_c is the protons repelling each other. The weak force does not enter the binding; mass differences decide which isotope transforms by beta decay.

Findings (the 2,457 measured nuclei of AME2020, Z, N ≥ 8):

Experiment: multiplying the Coulomb term by k (strong-force terms fixed; a thought experiment within the model):

In the real world the iron peak is Fe-56 / Ni-62, alpha decay is possible from A ≈ 145, and the theoretical estimate for Z ≤ 120 is 6,900 ± 500 bound nuclei (Erler et al. 2012); the model agrees with these. If the electromagnetic force were twice as strong, the most stable nucleus would shift to around silicon and the number of bound nuclei would fall to a third. Limit: if the fine-structure constant really changed, other quantities such as the neutron–proton mass difference would change too; the experiment keeps them fixed.

As a distributed campaign: the formula above takes seconds and needs no computing power. The version suited to mining is **a mass table for the whole chart of nuclei with nuclear density functionals (Skyrme / HFB) and an uncertainty analysis**: every (nucleus, parameter sample) is an independent calculation taking minutes to hours. The aim is to predict, with confidence intervals, the masses of neutron-rich nuclei (the r-process: how heavy elements such as gold form in stars). Open solvers (HFBTHO, Sky3D) should be evaluated (licences to be verified).

Interactive page: araclar/kuvvet_laboratuvari.html (opens with a double-click in the browser, no server needed). It scales the three forces with sliders and recomputes the chart of nuclei (Z ≤ 130), the binding curve, the measurement − model difference and the two-nucleon system.

Two nucleons (araclar/iki_nukleon.py; Yukawa (pion exchange) + Coulomb, s-wave, Numerov): range a = ħc / m_π c² = 1.414 fm. V₀ was fitted to two measurements: triplet channel 49.80 MeV (deuteron binding energy 2.2246 MeV), singlet channel 32.20 MeV (neutron–proton scattering length −23.74 fm).

Weak-force thresholds, with the neutron–proton mass difference Δ (1.293 MeV in reality):

• the free neutron decays: Δ > m_e (0.511 MeV),
• the hydrogen atom is stable (it does not capture its electron and turn into a neutron): Δ > m_e − 13.6 eV,
• p + p → d + e⁺ + ν in the Sun is possible: B_d − Δ − m_e > 0. In reality 0.420 MeV; had Δ exceeded 1.714 MeV, the Sun could not burn this way.

Had the strong force been 14% stronger, two neutrons would stay bound, and at 28% stronger two protons; had it been 27% weaker, there would be no deuteron, and so no first link in the chain that leads to all the heavy elements.

Scope limit: nuclear structure, decay energies and stellar reactions. Neutron transport, chain reactions, criticality and isotope enrichment calculations are deliberately left out of this layer.

8g. ORION-Compound: the thresholds at which elements form compounds

A campaign that maps, in two steps, which elements form stable compounds and in which ratios: a cheap formula filters the candidates, and miners compute the rest with DFT.

Tools:

• araclar/bilesik_laboratuvari.html: an interactive periodic table (similar to ptable.com's compound view). Pick elements, and the possible compounds, thresholds, estimated properties, known compounds (Wikidata) and DFT phases (Materials Project) are listed side by side.
• araclar/bilesik_modeli.py: downloads the data, fits the coefficients, validates the model and embeds the data in the page.

python araclar\bilesik_modeli.py indir ham                    # source data (~12 MB)
python araclar\bilesik_modeli.py veri ham veri.json            # page data + coefficient fit
python araclar\bilesik_modeli.py dogrula veri.json             # the validation table below
python araclar\bilesik_modeli.py tahmin veri.json Fe O         # candidates for one selection
python araclar\bilesik_modeli.py gom veri.json araclar\bilesik_laboratuvari.html

Thresholds (each is the precondition of the next):

The coefficients were fitted to experimental formation enthalpies (the Kingsbury compilation: Kim, Kubaschewski, NIST-JANAF; 798 charge-balanced compounds and 751 metal pairs): k = 0.681 eV (Pauling 1), c_N = 2.37 eV (Pauling 2.40), c_O = −0.08 eV (Pauling 1.05), k_m = 0.670 eV.

Predicted properties: a Turkish systematic name (demir(III) oksit = iron(III) oxide, diazot pentoksit = dinitrogen pentoxide, amonyum nitrat = ammonium nitrate; 33 polyatomic anion and ammonium templates; the lab names compounds in Turkish for now), bond type and ionic character, the shape and polarity of a single molecule (VSEPR; NH₃ trigonal pyramidal, SF₄ seesaw, XeF₄ square planar, ClO₂ radical), ionic lattice energy (Kapustinskii; NaCl 746, MgO 3,799, CaF₂ 2,638 kJ/mol, 0–8% from Born–Haber values), the temperature of decomposition into the elements. For known DFT phases: electrical class, Vickers hardness (Chen model, from elastic moduli) and magnetic moment.

Validation (dogrula, data of 2026-10-26):

Findings:

• Charge balance is the threshold of electrical behaviour: 81–92% of charge-balanced compounds are semiconductors or insulators, while 65% of unbalanced ones and 96% of metal pairs are metallic.
• Fitted to experiment, Pauling's form rediscovers the nitrogen correction by itself (2.37 eV versus Pauling's 2.40): the extra stability of the triple bond in N₂ shows clearly in the data. The oxygen correction is instead absorbed into a smaller k.
• The energy formula gives only a rough ranking. Because the Pauling sum is proportional to the number of bonds, it gives every oxide of the same metal the same energy per O: it cannot tell FeO from Fe₂O₃. 60% of the stable (metal + non-metal) binaries in MP are not generated: mixed-valence (Fe₃O₄), metal-rich (Ni₃P), with polyatomic anions (S₂²⁻ in FeS₂) or phases with oxidation states not in the list. The page shows these phases separately as "not generated by the model".
• The Ellingham threshold has the right magnitude (HgO decomposes at ~500 °C, the model says 612 °C). For Ag₂O the deviation is large (30 °C, in reality above ~200 °C): the model counts the entropy of solids as zero, which gives −103 instead of −66 J/mol·K for Ag₂O.

As a distributed campaign: the formulas take milliseconds; they are not worth mining, they filter candidates for free. The work distributed to miners is a DFT calculation for every (composition, structure) that passes the filter: the step that brings the formula's 0.42 eV/atom error down to 0.12.

9. The mining formula

The rules of sections 1 and 5 in numerical form. Notation: h block height; e epoch (E = 720 blocks ≈ 1 day); β = BILIM_PAYI; every amount is rounded down to the smallest unit (10⁻⁸ coin).

9a. Reward split

R(h)        = 10 coin · 2^(−⌊h / 1,050,000⌋)          block subsidy (zorluk.cpp: blok_odulu)
F_h         = transaction fees in the block
P_h         = ⌊β · R(h)⌋                               science pool (keyless pool output)
A_h         = ⌊0.05 · R(h)⌋ + ⌊0.20 · F_h⌋              archive pool (TASARIM.md section 11)
Miner_h     = R(h) + F_h − P_h − A_h

With β = 0.5 the miner receives 45% of the subsidy and 80% of the fees. With no campaign, P_h = 0.

No share is taken from fees. If half of the fee went to the pool, a miner could tell users "pay me the fee with a separate output"; the chain cannot tell the difference and the share would leak away. The subsidy, however, arises only from the consensus rule and cannot leak.

The pool balance in every block:

H_h = H_(h−1) + P_h + K_h − Ö_h          K_h: slashed stakes, Ö_h: science payments made

Slashed stakes go to the pool, not to the miner: the miner should have no interest in producing penalties.

9b. Epoch budget and payment

B_e     = ⌊λ · H_e⌋,  λ = 1/30                     amount distributed in the epoch (H_e: balance at the start of the epoch)
π_u     = w_c · c_u                                 points for a work unit
Π_j     = Σ π_u   (units j took part in where j agreed with the majority)
Ö_j     = ⌊B_e · Π_j / Σ_i Π_i⌋                      j's payment for the epoch

• c_u: the unit's reference cost, fixed per job type in the campaign manifest; core-seconds measured on a reference machine. The participant's actual time is not used and cannot be declared.
• w_c: campaign weight (science board, in the manifest). A core-second on a priority campaign is worth more; participants move there.
• 1/30 of the pool is distributed every day. Payments stay smooth even when participation fluctuates; in the long run, distributed = paid in. The equilibrium balance is H* = 30 × daily inflow. With β = 0.5 the daily inflow is 720 × 5 = 3,600 coins, H* ≈ 108,000 coins.
• Each of the k people who agree receives the unit's full points: everyone did the same computation. The cost of redundancy is that payment per core-second is divided by k.

9c. Assignment (lottery)

Unit u opens at height h_u. When the next block is found:

σ_u      = ORION-512(block_hash(h_u + 1) || "bilim-atama" || u)
U_i      = ORION-512(σ_u || address_i) / 2^512           uniform on (0, 1)
a_i      = S_i · ρ_i                                     stake × reputation
key_i    = −ln(U_i) / a_i
Chosen   = the k participants with the smallest keys     (Efraimidis–Spirakis weighted sampling)

• Because the block hash is unknown when the unit opens, no one can choose which unit they get. The miner who finds h_u + 1 would have to sacrifice the whole block to change the hash; since the only gain would be getting a job, it is not worth it.
• The probability of being chosen is proportional to a_i. Splitting stake across many identities does not raise the total probability; it lowers it, because new identities start with low reputation (ρ = 0.5). ρ rises with every agreeing unit (up to 1) and halves on disagreement.

9d. Commit, reveal, decision

commit_j = ORION-512(r_j || n_j || address_j)     r_j: canonical result, n_j: random 32 bytes

• Commit window T_t = 30 blocks (~1 hour), reveal window T_a = 30 blocks. All of a participant's commitments go, once an hour, as one Merkle root in a single transaction (section 5).
• Numerical results are rounded to the grid in the manifest (e.g. energy 1 meV/atom); **agreement = equality of the hash of the canonical result**. Nodes do not run scientific software; they only compare hashes.
• A unit is accepted if there are at least ⌊k/2⌋ + 1 equal results. If there is no majority, the unit is reopened with k new people and no one is penalized: honest computations that land on a grid boundary can also diverge.

9e. Penalties and audits

no reveal                       K = 0.01 · S_j
result against the majority     K = 0.10 · S_j    (if the majority ≥ 2 and the audit confirms the majority)
audit                           q = 5% of accepted units are chosen from σ and recomputed by k_d = 3 new people

A participant cheating alone (making up results without computing) is always caught: the other k − 1 honest results agree with each other. The real threat is collusion: a group holding a share f of the weight giving the same fake result. The probability of escaping detection:

P_escape = P_m(f, k) · (1 − q · (1 − P_m(f, k_d))),    P_m(f, k) = Σ_{m > k/2} C(k, m) f^m (1 − f)^(k−m)

Honesty condition. An honest participant earns v_u − κ·c_u from a unit; one who makes up results without computing earns P_escape·v_u − (1 − P_escape)·0.10·S_j (v_u: the unit's payment, κ: the electricity cost of one core-second):

cheating does not pay  ⇔  (1 − P_escape) · (v_u + 0.10 · S_j)  >  κ · c_u

As long as the payment covers the computing cost (v_u > κ·c_u / (1 − P_escape)), cheating for profit does not pay even without a stake. The stake's real job is to make an attacker who is not after profit but wants to corrupt the data pay: the expected cost of slipping one fake unit through is (1 − P_escape) / P_escape · 0.10 · S_j. Example: with f = 0.1, k = 3, S_j = 100·v_u, every fake unit that gets through costs ~360·v_u. Against a stronger attacker k = 5 is used, and results are also audited by the science board's random sampling before publication. Cheating in the science layer does not endanger the chain; it corrupts only part of the data and can be undone with a manifest version.

9f. Equilibrium: "half of the computing power"

A participant can give each core-second to proof of work or to science. As people shift toward the side that pays more, earnings per unit of computing power equalize on both sides:

Miner_h / X_PoW = P_h / X_science    ⇒    X_science / X_PoW = P_h / Miner_h = β·R / ((0.95 − β)·R + 0.80·F)

While fees are small, half of the computing power goes to science by itself; as fees grow, the balance shifts toward proof of work. Earnings set this balance, not a rule; the "Science %" slider in the app is a person's choice on their own computer. Limits: graphics cards and processors differ in efficiency at the two jobs (ORION is CPU-friendly, and DFT is mostly CPU-bound too); since the cost of an attack depends on miner revenue, taking over the chain gets cheaper as β grows (section 1).

Sources

• ZINC-22: PMC9976280, ZINC22 wiki
• Enamine REAL Space: BioSolveIT, September 2025 update, REAL Space
• VirtualFlow: Nature 2020 (PubMed), VirtualFlow 2.0 preprint
• Uni-Dock: GitHub, paper
• AutoDock Vina: GitHub · GNINA: GitHub, GNINA 1.0
• WHO BPPL 2024: WHO
• CO-ADD: open database, institution profile
• TDR Targets: about
• CARD: card.mcmaster.ca, licence
• ChEMBL: Bioregistry
• AlphaFold DB: NAR 2024
• OMol25: arXiv 2505.08762 · OMat24: arXiv 2410.12771
• OC20 / OC22: OC20, OC22
• ODAC25: ColabFit
• fairchem: GitHub · UMA: arXiv 2506.23971
• MACE: PyPI mace-torch
• Materials Project: re3data
• LIBE: Scientific Data 2021
• CoRE MOF 2024: Zenodo
• BOINC: BOINC: A Platform for Volunteer Computing
• Ofelimos: ePrint 2021/1379, FRLS: ePrint 2025/2091
• AME2020 atomic mass table: IAEA AMDC (Wang et al., Chinese Physics C 45, 030003, 2021)
• Bound-state beta decay of bare Dy-163: Jung et al., Physical Review Letters 69, 2164 (1992)
• Bound-state beta decay of bare Re-187: Bosch et al., Physical Review Letters 77, 5190 (1996)
• The limits of bound nuclei: Erler et al., "The limits of the nuclear landscape", Nature 486, 509 (2012)
• Formula for the total electron binding energy of atoms: Lunney, Pearson, Thibault, Reviews of Modern Physics 75, 1021 (2003)
• Charge balance + electronegativity screening rule (SMACT): Davies et al., "Computational screening of all stoichiometric inorganic materials", Chem 1, 617 (2016)
• Bond energy and ionic character: L. Pauling, The Nature of the Chemical Bond, 3rd edition, Cornell University Press (1960)
• Lattice energy: A. F. Kapustinskii, Quarterly Reviews, Chemical Society 10, 283 (1956) · Ionic radii: R. D. Shannon, Acta Crystallographica A32, 751 (1976)
• Hardness model: Chen, Niu, Li, Li, Intermetallics 19, 1275 (2011) · Ellingham diagram: H. J. T. Ellingham, J. Soc. Chem. Ind. 63, 125 (1944)
• Experimental formation enthalpies: Kim, Meschel, Nash, Chen, Scientific Data 4, 170162 (2017); Kingsbury compilation: Wang et al., ChemRxiv 14593476 (figshare, CC BY 4.0)
• Experimental band gaps: Zhuo, Mansouri Tehrani, Brgoch, J. Phys. Chem. Lett. 9, 1668 (2018) (figshare, CC BY 4.0)
• Materials Project: Jain et al., APL Materials 1, 011002 (2013); the 2018 snapshot via matminer (Ward et al., Comput. Mater. Sci. 152, 60, 2018)
• pymatgen resource files (MIT): GitHub (Ong et al., Comput. Mater. Sci. 68, 314, 2013)
• Element properties: PubChem periodic table · Known compounds and Turkish names: Wikidata (CC0), via the QLever endpoint
• Weighted lottery: Efraimidis, Spirakis, "Weighted random sampling with a reservoir", Information Processing Letters 97, 181 (2006)