Adamas in Chemistry, ML, and Axion Searches

• Adamas is a multi-disciplinary term that in diamondoid chemistry refers to adamantane—the smallest, highly symmetric diamondoid—while also naming distinct methods in machine learning and axion detection.
• In chemistry and materials science, adamantane serves as a benchmark molecule known for its structural rigidity, precise self-assembly behavior, and potential for nanodiamond synthesis.
• In machine learning, the Adamas method employs Hadamard transforms with dynamic token-level selection to achieve efficient sparse attention, in contrast to the unrelated ADAMOS axion haloscope.

Searching arXiv for the specific "Adamas" usages and adjacent literature. In recent arXiv literature, Adamas is not a single domain-stable term. In diamondoid chemistry, it is explicitly interpreted as adamantane (ADM), the prototypical lower diamondoid used as a reference system for self-assembly, spectroscopy, functionalization, and nanodiamond synthesis (Xue et al., 2018). In machine learning, Adamas denotes a training-free sparse attention method for long-context LLM inference (Yan et al., 21 Oct 2025). The term must also be distinguished from ADAMOS, the acronym for Axion DAily MOdulation Searches, a proposed fixed-frequency axion haloscope (Maroudas et al., 17 Feb 2026). This domain dependence is central to any rigorous use of the name.

1. Terminological scope and disambiguation

Within diamondoid research, adamantane is treated as the chemically primary referent for “Adamas”/ADM. It is described as the smallest diamondoid, a rigid hydrocarbon cage with high symmetry and a strain-free cage structure, and it functions as the benchmark against which diamantane and functionalized derivatives are compared in self-assembly and materials studies (Garcia et al., 2012, Xue et al., 2018). In that literature, the term points to a molecular scaffold rather than to a generic concept.

A distinct usage appears in long-context inference. There, Adamas is the name of a method that combines a Hadamard transform, bucketization, 2-bit compression, and Manhattan-distance estimation to perform dynamic token-level top-$k$ sparse attention during autoregressive decoding (Yan et al., 21 Oct 2025). The shared spelling does not imply any scientific relation to diamondoids.

A further source of confusion is ADAMOS, which expands to Axion DAily MOdulation Searches and refers to a proposed axion haloscope operating near 20 GHz in a 14 T superconducting magnet (Maroudas et al., 17 Feb 2026). This suggests that the term “Adamas” is best understood as a context-dependent label whose meaning is fixed by disciplinary usage rather than by etymology.

2. Adamantane as the chemical and materials-science referent

As adamantane, the chemical referent of “Adamas” is the molecule $\mathrm{C_{10}H_{16}}$, described as a rigid sp$^3$-bonded carbon cage and as the smallest diamondoid (Garcia et al., 2012). First-principles calculations reproduce its structural parameters closely: the C(1)–C(2) bond length is 1.538 Å, the average C–H bond length is about 1.105 Å, and the C–C–C bond angles are 109.5°, consistent with tetrahedral geometry (Garcia et al., 2012). The same study reports a HOMO–LUMO gap of 5.7 eV and a formation enthalpy of $-133.2$ kcal/mol, emphasizing exceptional molecular stability (Garcia et al., 2012).

That stability is paired with a limitation. Pristine adamantane is structurally robust, but its outer surface is hydrogen-terminated, so ordinary adamantane molecular crystals are held together mainly by weak dispersive forces and are correspondingly soft and brittle (Garcia et al., 2012, Garcia et al., 2012). This is why functionalization is central in the literature: the cage provides strong intramolecular rigidity, while substitutional chemistry is used to create designed intermolecular interactions (Garcia et al., 2012).

Several substitutions are treated as especially important. Boron- and nitrogen-substituted species include aza-adamantane, tetra-aza-adamantane, bora-adamantane, and tetra-bora-adamantane (Garcia et al., 2012). These are reported to remain energetically stable, with boron and nitrogen favored because they are close in size to carbon and introduce chemically active valence configurations without destroying the cage scaffold (Garcia et al., 2012). A key mechanistic point is that nitrogen contributes a lone-pair/nonbonding state, whereas boron provides an electron-deficient acceptor site, enabling strong directional intermolecular B$\leftarrow$N bonding (Garcia et al., 2012, Garcia et al., 2012).

3. Self-assembly, phase transitions, and crystal engineering

The self-assembly literature treats adamantane as the reference lower diamondoid against which derivatives are measured. Combined DFT and MD studies built seven separate simulation systems, each containing 125 molecules, and used a simulated-annealing protocol in which temperature was lowered in 1 K increments, each step lasting 10 ps with 5,000 time steps of 0.002 ps (Xue et al., 2018, Xue et al., 2018). The generic trajectory is reported as vapor $\rightarrow$ intermediate self-assembly $\rightarrow$ completed solid-like self-assembly (Xue et al., 2018).

The most ordered assemblies occur for adamantane and adamantane+Na, which show the sharpest radial distribution function peaks and the most distinct crystalline packing (Xue et al., 2018, Xue et al., 2018). Diamantane and diamantane+Na also form solid-like structures, but the order is less neat. By contrast, amantadine, memantine, and rimantadine form condensed but less ordered, non-crystalline assemblies. The structural reason given is that $-NH_2$ and related substituents disrupt symmetry and introduce hydrogen bonding whose spatial distribution is random, especially evident at 50 K and in hydrogen-bond angle statistics (Xue et al., 2018, Xue et al., 2018).

Density and substitution both shift transition behavior. For adamantane ensembles of 64 and 125 molecules, simulated at 5, 10, 20, 25, and 40 g/L, the reported trend is that higher density leads to higher phase transition temperatures, with marked finite-size differences at low density (Xue et al., 2018). Derivatives generally exhibit higher aggregation temperatures than the parent diamondoids, but the structural outcomes differ: amino-substituted derivatives aggregate at higher temperatures while losing crystal perfection, whereas Na-substituted derivatives raise the transition temperature while retaining much of the crystalline order (Xue et al., 2018).

Crystal-engineering studies extend this logic from self-assembly to designed solids. Functionalized adamantane molecules such as tetra-aza-adamantane (TA), tetra-bora-adamantane (TB), and di-aza-di-bora-adamantane (DADB) are arranged into zincblende and wurtzite architectures using first-principles calculations (Garcia et al., 2012). These molecular crystals are reported to have cohesive energies large enough for stability, bulk moduli of 20–42 GPa for functionalized systems, and 69–72 GPa for radical tetra-adamantyl crystals, with wide and direct band gaps of 3.8–4.4 eV and low dielectric constants $\kappa = 2.8$–$3.0$ (Garcia et al., 2012). A complementary study of the zincblende crystal formed from tetra-bora-adamantane and tetra-aza-adamantane reports a direct bandgap of 3.9 eV and a bulk modulus of 20 GPa, reinforcing the view that functionalized adamantanes can act as fundamental building blocks for nanostructure self-assembly (Garcia et al., 2012).

4. Spectroscopy, surfaces, and conversion to ultrasmall nanodiamonds

Surface-sensitive spectroscopy shows that adamantane’s vibrational response changes strongly when it is no longer isolated. For an adamantane monolayer on Au(111), measured by infrared scanning tunneling microscopy (IRSTM) and analyzed by DFT/DFPT, the monolayer forms hexagonally packed islands with a lattice constant of $\mathrm{C_{10}H_{16}}$0 Å, consistent with a $\mathrm{C_{10}H_{16}}$1 arrangement (Sakai et al., 2013). Two separate mechanisms modify the infrared spectrum relative to gas-phase adamantane: adamantane–adamantane packing reduces the 2912 cm$\mathrm{C_{10}H_{16}}$2 peak intensity by a factor of 3.5, while adamantane–gold interaction increases the 2938 cm$\mathrm{C_{10}H_{16}}$3 peak intensity by a factor of 2.6 and shifts it downward by 276 cm$\mathrm{C_{10}H_{16}}$4 (Sakai et al., 2013). The authors interpret the large redshift as a consequence of reduced electron density in the bottom C–H bonds caused by molecule–surface coupling.

Adamantane also serves as a direct precursor to diamond at extreme conditions. A high-pressure, high-temperature synthesis in a toroid-type high-pressure cell at 12 GPa and about 1300 °C converts adamantane into ultrasmall nanodiamonds with characteristic sizes of 2–5 nm (Kudryavtsev et al., 2022). Their Raman spectrum contains the downshifted diamond phonon near 1328 cm$\mathrm{C_{10}H_{16}}$5 and a distinctive broad band in the 1000–1500 cm$\mathrm{C_{10}H_{16}}$6 region with maxima at approximately 1147, 1245, 1344, and 1456 cm$\mathrm{C_{10}H_{16}}$7 (Kudryavtsev et al., 2022). These modes are assigned to CH$\mathrm{C_{10}H_{16}}$8 bending vibrations of hydrogen-terminated surface groups, and the especially intense 1344 cm$\mathrm{C_{10}H_{16}}$9 feature is explained by coupling with the 1328 cm$^3$0 diamond phonon (Kudryavtsev et al., 2022).

A significant interpretive claim in that work is that the unusual Raman band does not disperse with excitation wavelength, which rules out trans-polyacetylene and related assignments used in other nanodiamond contexts (Kudryavtsev et al., 2022). The proposed alternative is that the spectrum reflects the exceptionally high surface fraction and strong hydrogen termination of 2–5 nm particles produced from a hydrogen-rich precursor. The band is therefore proposed as an express, non-destructive way to recognize ultrasmall nanodiamonds synthesized from adamantane and related hydrogen-rich hydrocarbons, and the environmental sensitivity of polarized surface CH bonds is suggested to enable nanosensors in biology, chemistry, and medicine (Kudryavtsev et al., 2022).

5. Functional derivatives, optical emission, hydrogen storage, and aggregate energetics

Adamantane derivatives are also investigated as active functional materials. A notable example is single-crystalline 1,3,5,7-tetrakis-($^3$1-methoxyphenyl)adamantane, which exhibits octave-spanning emission across the visible spectrum under 325 nm excitation with a full width at half maximum of about 1.5 eV (Müller et al., 2022). The paper reports a photoluminescence quantum efficiency of about 45% at room temperature, an increase in emission area by about 40% on heating above room temperature, and external quantum efficiency above 7% at temperatures beyond 200 °C, with optical emission persisting up to 475 K (Müller et al., 2022). The mechanistic interpretation is based on self-trapped excitons (STEs): a central broad band is assigned to bulk self-trapped excitons, a higher-energy component to surface photoluminescence, and a lower-energy component to deeper or tail states (Müller et al., 2022).

Another functional direction is hydrogen storage. First-principles work on Li and Li$^3$2 functionalized adamantane shows that replacing an acidic hydrogen activates the molecule for molecular hydrogen adsorption (Ranjbar et al., 2011). Each Li or Li$^3$3 site can bind up to five $^3$4 molecules, with reported binding energies in the range $^3$5 to $^3$6 eV/$^3$7 for ADM.Li and $^3$8 to $^3$9 eV/$-133.2$0 for ADM.Li$-133.2$1 (Ranjbar et al., 2011). The estimated gravimetric hydrogen-storage capacity is around $-133.2$2 wt% for ADM.Li$-133.2$3, and the mechanism is described as electrostatic polarization of $-133.2$4 by the electric field of positively charged Li/Li$-133.2$5 rather than dissociative chemisorption (Ranjbar et al., 2011). The paper also reports that ADM.Li can aggregate, whereas ADM.Li$-133.2$6 does not show clustering because of Coulomb repulsion (Ranjbar et al., 2011).

At a more abstract structural level, computed studies of diamondoid “polymer-like” aggregates built from adamantane and diamantane cages show regular scaling laws in both PM6 total energies and MM2 SWB-tension energies (Balaban et al., 2015). For five families—spiro-$-133.2$7adamantane, spiro-$-133.2$8diamantane, one-bond-sharing-$-133.2$9adamantane, one-bond-sharing-$\leftarrow$0diamantane, and 1234-helical-cata-$\leftarrow$1diamantanes—acyclic SWB-tension energies increase essentially linearly with $\leftarrow$2, while cyclic aggregates typically pass through a minimum at an intermediate $\leftarrow$3 (Balaban et al., 2015). A central conclusion is that cyclic and acyclic sequences of a given class approach a common limiting energy per unit as $\leftarrow$4, which the paper interprets as a bulk-like asymptotic limit (Balaban et al., 2015).

6. Adamas in long-context inference

In machine learning, Adamas is a training-free sparse attention method for long-context LLM inference (Yan et al., 21 Oct 2025). The method starts from the standard attention operator,

$\leftarrow$5

and addresses the quadratic scaling of $\leftarrow$6 over long contexts by avoiding exhaustive query–key scoring during decoding (Yan et al., 21 Oct 2025). Its central mechanism is a two-stage procedure: first, approximate retrieval of relevant tokens using compressed transformed keys; second, exact sparse attention on the selected subset.

The transformation stage applies a Hadamard transform to both queries and keys,

$\leftarrow$7

using the orthogonality relation $\leftarrow$8 to preserve dot-product similarity in principle (Yan et al., 21 Oct 2025). After transformation, coordinates are bucketized into four levels $\leftarrow$9, encoded with 2 bits per value, and packed so that every 8 elements occupy a 16-bit value, adding only about $\rightarrow$0 extra cache cost (Yan et al., 21 Oct 2025). Similarity is then estimated by the negative Manhattan distance between compressed vectors,

$\rightarrow$1

followed by dynamic token-level top-$\rightarrow$2 selection and exact attention over the retrieved original keys and values (Yan et al., 21 Oct 2025).

The method is explicitly positioned against two failure modes of prior sparse attention: static sparse patterns that are not query-aware, and dynamic page-level retrieval that remains too coarse to recover scattered critical tokens reliably (Yan et al., 21 Oct 2025). Its ablation study identifies the Hadamard transform as essential: direct low-bit bucketization without it performs very poorly at small budgets, which the paper attributes to outliers in the raw query/key coordinates (Yan et al., 21 Oct 2025). The comparison of 1-bit, 2-bit, and 3-bit bucketization selects 2-bit as the best trade-off, while the comparison of Manhattan and L2 distance finds broadly similar accuracy with L1/Manhattan judged more robust to noise and sparsity (Yan et al., 21 Oct 2025).

Empirically, the reported results are strong. On long-context benchmarks, Adamas is said to match the accuracy of full attention with only a 64-token budget, achieve near-lossless performance at 128, and support up to 8x higher sparsity than prior state of the art while delivering up to 4.4x self-attention and 1.5x end-to-end speedups on 32K-length sequences (Yan et al., 21 Oct 2025). On passkey retrieval with 10K-length inputs, the method reports 68% at budget 16, 85% at 32, 93% at 64, and near saturation above 128. On 100K-length inputs, it reports 54% at budget 64, 71% at 128, and 87% at 256 (Yan et al., 21 Oct 2025). On PG19, the paper further claims that Adamas can attain comparable or even lower perplexity than full attention, suggesting that sparse retrieval may sometimes suppress irrelevant context rather than merely approximate dense attention (Yan et al., 21 Oct 2025).

7. Distinction from ADAMOS and the broader significance of the name

A third technical usage, orthographically close but semantically unrelated, is ADAMOS, the proposed Axion DAily MOdulation Searches experiment (Maroudas et al., 17 Feb 2026). ADAMOS is a fixed-frequency cavity resonator near 19.95 GHz, corresponding to an axion mass near 82.7 μeV, designed around a thin-shell cavity formed by two concentric OFHC copper cylinders separated by an annular gap of about 7.5 mm (Maroudas et al., 17 Feb 2026). The geometry has an outer diameter of 125 mm, an inner-cylinder outer diameter of 94 mm, a length of 400 mm, and a total volume of 0.96 liters, yielding a pseudo-TM$\rightarrow$3 mode with approximately 25 times the detection volume of a conventional cavity at the same frequency (Maroudas et al., 17 Feb 2026).

The experiment is to operate in an existing 14 T warm-bore superconducting magnet and is designed to search simultaneously for conventional cold dark matter axions, relativistic axions from axion quark nugget annihilations exhibiting daily modulation, and transient enhancements from streaming dark matter (Maroudas et al., 17 Feb 2026). Its simulated cavity performance is summarized by $\rightarrow$4, $\rightarrow$5, and $\rightarrow$6, and the projected sensitivity after 30 days at 19.95 GHz is $\rightarrow$7 for $\rightarrow$8 (Maroudas et al., 17 Feb 2026). For the daily-modulation channel, the estimated Hamburg modulation amplitude is about 5.4%, motivating continuous in situ calibration to control temperature-dependent gain drifts (Maroudas et al., 17 Feb 2026).

Taken together, these literatures indicate that “Adamas” is not a unified scientific object but a domain-indexed label. In chemistry and materials science it maps to adamantane/ADM and its derivatives; in machine learning it names a Hadamard sparse attention mechanism; and it must be kept separate from ADAMOS in axion detection. A plausible implication is that precise disciplinary qualification is necessary whenever the term appears in technical writing.