SAND2025-12355R SANDIA REPORT SAND2025-12355R Printed October 2025 Probability Density Function for the spatial and intensity distribution of neutron-induced defects in Silicon Daniel M. Fletcher, Brian D. Hehr, and Nicholas A. Asper Prepared by Sandia National Laboratories Albuquerque, New Mexico 87185 Livermore, California 94550 CONTENTS Acknowledgement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2. Computational Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.1. Molecular Dynamics: LAMMPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2. BCA: SRIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.3. BCA: MARLOWE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.4. Radiation Transport: MRED and G4SEE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3. Intensity Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.1. Mono-energetic Projectiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.1.1. MARLOWE Modeling Sensitivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.2. Calibrating MARLOWE to MD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.2.1. Molecular Dynamics Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.2.2. Calibrating Displacements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.2.3. Calibrating Stable Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.2.4. Calibration Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.3. Distribution Fits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.3.1. Alternate Fits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.4. Recoil Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.4.1. Sampling Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 3.4.2. Probability Density Functions for Incident Neutrons . . . . . . . . . . . . . . . . . . . . 50 4. Spatial Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.1. FinFET Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.2. Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 4.2.1. Multi-channel Defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 4.2.2. Comparisons with SRIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 4.2.3. Recoil Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 5.1. Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 A. Multi-layer MARLOWE input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 7 B. Multi-layer TRIM input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 8 LIST OF FIGURES Figure 3-1. Comparing probability densities for the number of Frenkel pairs using the crystalline and amorphous composition models . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 Figure 3-2. Comparing probability densities for the fraction of the recoil energy lost to in- elastic collisions (ionization) using the crystalline and amorphous composition models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Figure 3-3. Comparing probability densities for the number of Frenkel pairs using the local and non-local inelastic energy loss models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Figure 3-4. Comparing probability densities for the fraction of the recoil energy lost to inelastic collisions (ionization) using the local and non-local inelastic energy loss models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Figure 3-5. Comparing probability densities for the number of Frenkel pairs using the local and non-local inelastic energy loss models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 Figure 3-6. Comparing probability densities for the fraction of the recoil energy lost to inelastic collisions (ionization) using the local and non-local inelastic energy loss models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Figure 3-7. Comparing average quantities predicted by the model combinations . . . . . . . . . . 28 Figure 3-8. Comparing probability densities for the number of Frenkel pairs using various cutoff energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Figure 3-9. Comparing probability densities for the fraction of the recoil energy lost to inelastic collisions (ionization) using various cutoff energies . . . . . . . . . . . . . . . . 30 Figure 3-10. Average Frenkel pair production and inelastic energy loss fraction for various cutoff energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Figure 3-11. Example of crystalline silicon lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 Figure 3-12. ZBL fraction for the silicon Tersoff interatomic potential . . . . . . . . . . . . . . . . . . . 33 Figure 3-13. Cohesive Energy of silicon as a function of lattice constant . . . . . . . . . . . . . . . . . 34 Figure 3-14. Example of a cascade in silicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Figure 3-15. Number of displacements vs time for various initial PKA events . . . . . . . . . . . . . 35 Figure 3-16. Example of the Wigner-Seitz method for a 5 keV cascade in silicon . . . . . . . . . . 36 Figure 3-17. The effect of varying EBND(3) on the total number of displaced atoms predicted by MARLOWE in comparison to MD as a function of PKA energy . . . . . . . . . . 38 Figure 3-18. The effect of varying the recombination range on the total number of Frenkel pairs predicted by MARLOWE in comparison to MD as a function of PKA energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 Figure 3-19. Total Displaced Ions - Comparison of probability density functions calculated using calibrated MARLOWE and MD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 Figure 3-20. Stable Pairs - Comparison of probability density functions calculated using calibrated MARLOWE and MD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 9 Figure 3-21. Correlation between inelastic loss fraction and stable pair formation for various PKA energies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 Figure 3-22. The ten best-fit probability density functions for the number of stable Frenkel Pairs for two PKA energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 Figure 3-23. The ten best-fit probability density functions for the inelastic energy loss frac- tion for two PKA energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Figure 3-24. Fit parameters of the generalized logistic function for a range of recoil energies 46 Figure 3-25. Kernel Density Estimator fits (smooth) to MARLOWE results (histograms) . . . 47 Figure 3-26. TableCurve 2D fit to stable pair distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 Figure 3-27. Si recoil distributions for 1 and 14 MeV incident neutrons . . . . . . . . . . . . . . . . . . 49 Figure 3-28. Rejection Sampling vs. Weighted Uniform Samples . . . . . . . . . . . . . . . . . . . . . . . 50 Figure 3-29. Damage quantities for 1-MeV incident Neutrons . . . . . . . . . . . . . . . . . . . . . . . . . . 51 Figure 3-30. Damage quantities for 14-MeV incident Neutrons . . . . . . . . . . . . . . . . . . . . . . . . . 52 Figure 3-31. Probability Density Functions for 14 MeV incident Neutrons, comparisons with SRIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 Figure 4-1. Longitudinal and cross sectional dimensions of the GAAFET NS model. The highlighted region is modeled in MARLOWE . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 Figure 4-2. Example Cascades for various PKA energies. The plots show the final locations of the interstitial atoms of each species. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 Figure 4-3. 3D plot of the 5 keV cascade shown in Figure 4-2 . . . . . . . . . . . . . . . . . . . . . . . . . 58 Figure 4-4. Probability density functions for stable pair formation in multiple layers for a range of PKA energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 Figure 4-5. Probability density functions for stable pair formation in multiple layers for a range of PKA energies, comparison between MARLOWE, "SRIMized" MAR- LOWE, and SRIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 Figure 4-6. Probability density functions for stable pair formation in multiple layers for 10 keV recoil ions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 Figure 4-7. Probability density functions for stable pair formation in multiple layers for 1 and 14 MeV incident neutrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 10 LIST OF TABLES Table 3-1. Number of Frenkel pairs predicted by the crystalline and amorphous composition models in MARLOWE for a range of recoil ion energies . . . . . . . . . . . . . . . . . . . . . 20 Table 3-2. Inelastic energy loss fraction predicted by the crystalline and amorphous compo- sition models in MARLOWE for a range of recoil ion energies . . . . . . . . . . . . . . . . 21 Table 3-3. Number of Frenkel pairs predicted by the local and non-local inelastic energy loss models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Table 3-4. Inelastic energy loss fraction predicted by the local and non-local inelastic energy loss models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Table 3-5. MD-Calibrated Model Options for Si-Si calculations in MARLOWE . . . . . . . . . . . 39 11 12 ACRONYMS & DEFINITIONS BCA Binary Collision Approximation GAA NS Gate-All-Around Nanosheet KDE Kernel Density Estimator MD Molecular Dynamics MLE Maximum Likelihood Estimation NIEL Non-Ionizing Energy Loss PKA Primary Knock-On Atom RSICC Radiation Safety Information Computational Center SEE Single Event Effects SNL Sandia National Laboratories ZBL Ziegler-Biersack-Littmark 13 14 1. INTRODUCTION The proper functioning of electronic devices in radiation environments is a key concern within Sandia’s survivability mission. Several mechanisms of radiation-induced degradation of electronics exist, some of which are particularly sensitive to neutron environments which can produce heavy ion recoils that generate a high density of ionization in the immediate vicinity of the energetic recoil event. Furthermore, recoils from high-energy neutron interactions have a significant probability of being accompanied by lighter reaction products such as alpha particles, which can generate ionization at substantially farther ranges in comparison to the heavy residual ions. Energetic neutron interactions in materials give rise to displacement damage and ionization, which are known to be stochastic phenomena exhibiting a large relative standard deviation in relevant metrics such as Frenkel pair production and cluster size. In addition to this intrinsic variability, model parameterization can also strongly influence such metrics in a more systematic manner. Because displacement damage and ionization are both linked to the degradation of semiconductor devices, it is important to understand the fundamental nature of damage from irradiation as well as the likelihood of events that will induce upsets in larger-scale systems. State-of-the-art systems (particularly for non-volatile memory) often incorporate error correction mechanisms to recover from individual bit upsets. Since the recovery function requires some characteristic time to execute, it is possible to overwhelm the system with a sufficiently high rate of upsets such that contiguous memory elements upset in an effectively simultaneous and irrecoverable manner. This is termed a “multi-bit upset” (MBU) and, significantly, it may occur in lower flux conditions through single particle interactions that create damage over a large enough volume to upset multiple memory elements. Such “low probability, high consequence” occurrences must be quantified to make probabilistic assurances of system reliability. Computational tools such as the Transport of Ions in Matter (SRIM/TRIM) Binary Collision Approximation (BCA) code [1], the MARLOWE BCA code [2], and the LAMMPS molecular dynamics (MD) code [3] will be used to investigate the short-term damage production process from the time of neutron interaction until the formation of a quasi-stable array of defects—a time span on the order of picoseconds. While the BCA technique is limited to the immediate outcome of the recoil-induced collision cascade, MD simulations, in the modern computing environment, can realistically be extended out to nanoseconds or longer, but, for our modeling purposes of representing the initial Frenkel pair production and charge deposition, calculations out to picoseconds are adequate. “Damage” here encompasses the amount and spatial layout of ionization as well as point defects such as Frenkel pairs. Beyond the post-irradiation time at which the cascade-induced thermal spike has dissipated and thermally-activated defect migration (at ambient temperature) becomes necessary for further defect-defect or defect-impurity reactions, evolution of the damage typically must be handled using other techniques, such as finite element drift-diffusion or kinetic Monte Carlo (kMC). 15 16 2. COMPUTATIONAL MODELS 2.1. Molecular Dynamics: LAMMPS LAMMPS is a classical Molecular Dynamics (MD) code meant for modeling computationally material large structures in parallel. Within LAMMPS, many-body interatomic potentials are used to model the dynamic behavior of the material given an initial state and boundary condition. For energetic events, such as displacement cascades, a repulsive and chemical interatomic potential are hybridized for the simulation. An advantage that MD codes like LAMMPS has is that it can model the short-time annealing (1-100 picoseconds) of defects produced from a Primary Knock-On Atom (PKA) event. However, the computational costs of running MD scale O(n3 ) with the number of atoms and the minimum width of the simulation box scales proportionally the recoil energy of the PKA event. As a result, the energetic cap for MD simulations of displacement cascades, due to computational limitations, are on the order of 10s of keV for heavy ions. 2.2. BCA: SRIM SRIM is a freely-available and popular tool for performing fast-running simulations of energetic ion impingement on multi-layer material stackups. Collision cascade events are modelled under the binary collision approximation with a further assumption of amorphous media, which simplifies the characterization of collision events (and, of course, disallows phenomena particular to crystalline materials such as channeling and focusing). In spite of its widespread adoption within the radiation effects and ion implantation communities, the proper means of deploying SRIM for various applications has been subject to debate, and specific recommendations have been published on mode of execution and input parameter selection [4, 5]. A recurring topic of contention is the distinction between the “full cascade” and “quick” mode of TRIM, the former being a full Monte Carlo BCA treatment of PKAs and SKAs and the latter applying the Norgett-Robinson-Torrens (NRT) formalism to SKAs produced along the incident ion track. It is generally recognized that the full cascade mode invokes a more physically sound approach, especially in multi-elemental systems [5]. Quick mode, on the other hand, aligns closer with the theoretical underpinnings of certain NRT-based standard metrics such as radiation exposure (in units of displacements per atom (DPA)). As MARLOWE and SRIM utilize the same basic methodology of simulating displacement damage, most input parameters and outputted quantities are directly analogous. Significant differences exist, however. For example, both codes require lattice binding energies that correspond to the “energetic penalty” of defect production, but SRIM makes use of the threshold displacement energy whereas MARLOWE defines a “cutoff energy” (not necessarily equivalent to the threshold 17 displacement energy). Also, the ZBL interatomic potential is the only option permitted in SRIM while MARLOWE accommodates several potential functions. 2.3. BCA: MARLOWE MARLOWE is a BCA code which has an advantage over SRIM in that it is capable of treating the effect of channeling due to crystalline structure of the medium. While development on the official MARLOWE distribution available from Radiation Safety Information Computational Center (RSICC) appears to have ceased, it continues to be used[6] and developed[7, 8] independently. The base MARLOWE currently in use at Sandia National Laboratories (SNL) is version 15b from RSICC, which has undergone several modifications to add Ziegler-Biersack-Littmark (ZBL) inelastic loss model[9], increase numerical precision, increase the maximum cascade size, and fix numerous minor bugs. MARLOWE has numerous model options to treat the elastic and inelastic processes. 2.4. Radiation Transport: MRED and G4SEE The open-source G4SEE[10] was investigated for use in this work. G4SEE is a toolkit built upon the Monte Carlo transport code GEANT4, and is aimed at being a user-friendly interface to the physics options relevant to modeling Single Event Effects (SEE). MRED[11] is a similar tool developed at Vanderbilt University also built upon GEANT4. The advantage these codes have is that where a user must supply the characteristics of the PKA in a BCA code, the PKA can be sampled randomly according to the recoil distributions defined in the incident neutron nuclear data libraries. This vastly simplifies the problem definition where, for example, the PKA is produced by a neutron strike in one material, but deposits its energy in a different, adjacent material. These codes are specialized to the SEE application, and inputting appropriate physics, geometry and biasing options is simplified. A disadvantage to these codes is that they cannot model the recoil cascade in detail. While Non- Ionizing Energy Loss (NIEL) can be separated from the ionizing energy deposition along a particle track, it is not possible to extract details of the defect production in the crystal structure. There is a capability[12] to model the effect of particle channeling in a crystalline lattice in GEANT4, but it is not clear how a user would integrate those models in to the codes. G4SEE has an additional disadvantage with respect to the work here in that it can only model a single sensitive volume. Neither of these codes were used for the work here but could perhaps be used to to extend the work performed as part of this study generate detailed spectra of heavy and light recoil ions a produced in or entering a sensitive volume. 18 3. INTENSITY DISTRIBUTION The methods for deriving the probability density functions for Frenkel pair production and ion- ization in Silicon are described in this section. The modeling sensitivities inherent in the BCA approximation will be explored, before deriving a set of parameters and physics options that pro- duce probability density functions which are in good agreement with high-fidelity MD simulations. Finally, analytic probability distributions will be fitted to the Frenkel pair and ionization energy fractions to provide a means of estimating the probabilities of rare but high-consequence events. 3.1. Mono-energetic Projectiles Before considering the recoil energy distributions produced by an incident neutron, we first seek to characterize the probability density functions of single ion energies. Si28 PKAs are sampled randomly in space and injection angle. The initial azimuthal angle is sampled uniformly between 0 and 360 degrees. A cosine weighting is applied to the polar angle: 𝜃 = arccos(1 − 𝑈 (0, 2)) where U(0,2) is a uniform random variable defined in the range 0 to 2. 3.1.1. MARLOWE Modeling Sensitivities Aside from the binding energy and recombination distance, there are a number of model options in MARLOWE which have a non-negligible impact on the results. Here we will investigate: 1) the crystal model - amorphous vs crystalline, 2) the inelastic energy loss model - local vs non-local, and 3) the cut-off energy - the energy at which an ion trajectory is terminated. 3.1.1.1. Amorphous vs. Crystalline The POLY parameter within the &XTAL option group is used to select the crystallinity of a given layer. POLY=0 defines the layer as mono-crystalline, while POLY=1 defines the layer as poly-crystalline. For this option, the crystal lattice in each layer is randomly rotated at the start of each PKA trajectory. Since the PKAs are to be injected at random angles this option is unused here. Finally POLY=2 defines the layer as amorphous. Since the ion transport algorithm in MARLOWE is inherently based upon the physical locations of the lattice sites (as opposed to SRIM), an amorphous material 19 Table 3-1. Number of Frenkel pairs predicted by the crystalline and amorphous composition models in MARLOWE for a range of recoil ion energies Recoil Energy (keV) Crystalline Amorphous Mean Std. Dev. Mean Std. Dev. 2 45.6 6.9 (15.1 %) 59.8 5.9 (8.5 %) 5 102.8 16.2 (15.8 %) 134.5 11.7 (8.2 %) 10 189.5 34.0 (18.0 %) 246.5 21.3 (8.4 %) 50 739.3 452.1 (26.0 %) 929.9 99.5 (10.7 %) 100 1227.7 880.1 (30.2 %) 1550.8 193.8 (12.5 %) 500 3211.6 1246.6 (38.8 %) 3974.6 672.3 (16.96 %) is emulated by randomly rotating the lattice after each encounter having an impact parameter less than the mean atomic radius. In all these calculations, the cutoff energy is 1 eV, and the non-local inelastic energy loss model is used. Ten thousand primary atoms are simulated for each energy. The means and standard deviations for the number of Frenkel pairs and the inelastic energy loss for a variety of recoil ion energies are given in Tables 3-1 and 3-2, respectively. In terms of Frenkel pairs, the difference in the mean between the crystalline and amorphous composition is around 20% for all the ion energies. However, the standard deviation differs more drastically, with the crystalline model predicting a value 2-3 times that predicted by the amorphous model. A similar trend is observed for the inelastic energy losses. Figure 3-1 compares the probability densities for the number of Frenkel pairs for various recoil ion energies. It can be seen that not only does the standard deviation show large variation, but the distributions themselves differ significantly. The amorphous model predicts a somewhat Gaussian shape, while the crystalline model predicts a longer tail to the left of the distribution. The inelastic fraction is shown in Figure 3-2. This follows a similar trend to the displacement distributions, except the long tail extends to the right. This is an intuitive result considering the kinematics involved in each composition model. The inelastic energy loss can be considered to be constantly accumulating between elastic collisions. In the crystalline model, an ions trajectory may allow it to channel through the crystal relatively large distances without an elastic collision occurring with a lattice atom. In the extreme it is possible that the PKA loses all its energy to inelastic processes, in which case the inelastic fraction is 1 and the number of displacements is zero. Such trajectories are essentially impossible in the amorphous model. 3.1.1.2. Local vs. Non-local inelastic energy loss The ZBL potential (as implemented in MARLOWE) has two inelastic energy loss model options: local and non-local. In the non-local model, the inelastic energy loss is proportional to the energy 20 Table 3-2. Inelastic energy loss fraction predicted by the crystalline and amorphous composition models in MARLOWE for a range of recoil ion energies Recoil Energy (keV) Crystalline Amorphous Mean Std. Dev. Mean Std. Dev. 2 0.51 0.06 0.52 0.03 5 0.55 0.06 0.56 0.03 10 0.59 0.07 0.60 0.03 50 0.68 0.08 0.70 0.03 100 0.73 0.08 0.75 0.03 500 0.86 0.05 0.87 0.02 Table 3-3. Number of Frenkel pairs predicted by the local and non-local inelastic energy loss models Recoil Energy (keV) Non-Local Local Mean Std. Dev. Mean Std. Dev. 2 45.6 6.9 (15.1 %) 62.8 6.2 (9.8 %) 5 102.8 16.2 (15.9 %) 141.9 16.1 (11.3 %) 10 189.5 34.1 (18.0 %) 261.2 37.5 (14.3 %) 50 730.3 189.9 (26.0 %) 1008.6 233.0 (23.1 %) 100 1227.7 370.6 (30.2 %) 1707.6 470.2 (27.5 %) 500 3211.2 1246.6 (38.8 %) 4553.1 1679.6 (37.0 %) and the track length between elastic collisions, and can be considered a constant viscous force acting upon the ion: −𝑑𝐸 = 𝑍 𝐵𝐿 (𝐸) (3.1) 𝑑𝑥 where MARLOWE integrates −𝑑𝐸/𝑑𝑥 along the trajectory between collision sites. In the local model the inelastic energy loss is also a function of the impact parameter of the target atoms encountered along the trajectory. In the calculations presented in this section, the crystalline composition model is used, and once again, the energy cutoff is set to 1 eV. Tables 3-3 and 3-4 compare the means and standard deviations predicted by the local and non-local models. In this case, the standard deviations between the two models are similar; however there is a shift in the means. On average, the local model predicts around 30-40% more Frenkel pairs and 10-20% less energy lost to inelastic processes. The distributions are shown in Figures 3-3 and 3-4. In contrast to the comparison between the composition models, changing the inelastic energy loss model does not appear to significantly alter the shape of the distribution; instead, it shows only a shift in the peaks. 21 2 keV 5 keV 10 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 Crystalline Amorphous 10−6 10−6 10−6 0 50 0 100 0 200 50 keV 100 keV 500 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 10−6 10−6 10−6 0 500 1000 0 1000 2000 0 5000 Figure 3-1. Comparing probability densities for the number of Frenkel pairs using the crystalline and amorphous composition models All four combinations of composition and inelastic energy loss model were tested to investigate any potential synergistic effects. Figures 3-5 and 3-7, however do not show that to be the case, and lead to the same conclusions as before: the choice of inelastic loss model shifts the means, and the choice of composition model changes the distribution and standard deviation. The means and standard deviations for both quantities are compared in Figures 3-5 and 3-7. 3.1.1.3. Energy cut-off The energy cut-off (EQUIT in MARLOWE) is the energy at which an ion track is terminated. Once an ion reaches this cut-off, the remaining energy is reported as "subthreshold energy loss". There is a broad range of suggested cut-off energies in the literature. Some authors suggest 1- 2 eV[13] based on the limits of the applicability of the BCA model; others suggest using some multiple of the displacement energy[14] or the cohesive energy[15]. Authors employing Hybrid MD-BCA methods suggest cutoffs in excess of 200 eV[16, 17]. 22 2 keV 5 keV 10 keV 102 102 102 101 101 101 100 100 100 10−1 10−1 10−1 10−2 Crystalline 10−2 10−2 Amorphous 10−3 10−3 10−3 0.4 0.6 0.8 0.6 0.8 0.6 0.8 50 keV 100 keV 500 keV 102 102 102 101 101 101 100 100 100 10−1 10−1 10−1 10−2 10−2 10−2 10−3 10−3 10−3 0.6 0.8 1.0 0.6 0.8 1.0 0.8 1.0 Figure 3-2. Comparing probability densities for the fraction of the recoil energy lost to inelastic collisions (ionization) using the crystalline and amorphous composition models Here, we explore the sensitivity of the number of pairs and inelastic loss fraction to the cutoff energy for a range of PKA energies. The crystalline composition and the non-local inelastic energy loss models are used. 23 Table 3-4. Inelastic energy loss fraction predicted by the local and non-local inelastic energy loss models Recoil Energy (keV) Crystalline Amorphous Mean Std. Dev. Mean Std. Dev. 2 0.51 0.06 0.28 0.05 5 0.56 0.06 0.35 0.07 10 0.59 0.07 0.40 0.08 50 0.68 0.08 0.53 0.10 100 0.73 0.08 0.51 0.11 500 0.86 0.05 0.79 0.08 2 keV 5 keV 10 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 Non-Local Local 10−6 10−6 10−6 0 50 0 100 0 200 50 keV 100 keV 500 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 10−6 10−6 10−6 0 1000 0 1000 2000 0 5000 Figure 3-3. Comparing probability densities for the number of Frenkel pairs using the local and non-local inelastic energy loss models 24 2 keV 5 keV 10 keV 102 102 102 101 101 101 100 100 100 10−1 10−1 10−1 10−2 Non-Local 10−2 10−2 Local 10−3 10−3 10−3 0.25 0.50 0.75 0.25 0.50 0.75 0.25 0.50 0.75 50 keV 100 keV 500 keV 102 102 102 101 101 101 100 100 100 10−1 10−1 10−1 10−2 10−2 10−2 10−3 10−3 10−3 0.50 0.75 1.00 0.50 0.75 1.00 0.6 0.8 1.0 Figure 3-4. Comparing probability densities for the fraction of the recoil energy lost to inelastic collisions (ionization) using the local and non-local inelastic energy loss models 25 2 keV 5 keV 10 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 10−6 10−6 10−6 0 50 100 0 100 200 0 200 400 50 keV 100 keV 500 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 10−6 10−6 10−6 0 1000 0 2000 0 5000 10000 Crystalline - Non-Local Amorphous - Non-Local Crystalline - Local Amorphous - Local Figure 3-5. Comparing probability densities for the number of Frenkel pairs using the local and non-local inelastic energy loss models 26 2 keV 5 keV 10 keV 101 101 101 10−1 10−1 10−1 10−3 10−3 10−3 0.25 0.50 0.75 0.25 0.50 0.75 0.25 0.50 0.75 50 keV 100 keV 500 keV 101 101 101 10−1 10−1 10−1 10−3 10−3 10−3 0.50 0.75 1.00 0.50 0.75 1.00 0.6 0.8 1.0 Crystalline - Non-Local Amorphous - Non-Local Crystalline - Local Amorphous - Local Figure 3-6. Comparing probability densities for the fraction of the recoil energy lost to inelastic collisions (ionization) using the local and non-local inelastic energy loss models 27 103 102 Crystalline - Non-Local Crystalline - Local Amorphous - Non-Local Amorphous - Local 101 0 100 200 300 400 500 Recoil Energy [keV] (a) Average Frenkel pair production 0.9 0.8 0.7 0.6 0.5 0.4 Crystalline - Non-Local 0.3 Crystalline - Local Amorphous - Non-Local Amorphous - Local 0 100 200 300 400 500 Recoil Energy [keV] (b) Average inelastic energy loss fraction Figure 3-7. Comparing average quantities predicted by the model combinations 28 2 keV 5 keV 10 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 10−6 10−6 10−6 0 25 50 0 100 0 100 200 50 keV 100 keV 500 keV 100 100 100 10−2 10−2 10−2 10−4 10−4 10−4 10−6 10−6 10−6 0 500 1000 0 1000 2000 0 2500 5000 1 eV 5 eV 20 eV 2 eV 10 eV Figure 3-8. Comparing probability densities for the number of Frenkel pairs using various cutoff energies 29 2 keV 5 keV 10 keV 101 101 101 10−1 10−1 10−1 10−3 10−3 10−3 0.50 0.75 0.50 0.75 0.50 0.75 50 keV 100 keV 500 keV 101 101 101 10−1 10−1 10−1 10−3 10−3 10−3 0.6 0.8 1.0 0.6 0.8 1.0 0.8 1.0 1 eV 5 eV 20 eV 2 eV 10 eV Figure 3-9. Comparing probability densities for the fraction of the recoil energy lost to inelastic collisions (ionization) using various cutoff energies 30 1 eV 2 eV 4000 5 eV 10 eV 20 eV 3000 2000 1000 0 0 100 200 300 400 500 Recoil Energy [keV] (a) Average Frenkel pair production 0.9 0.8 0.7 0.6 0.5 1 eV 2 eV 0.4 5 eV 10 eV 20 eV 0 100 200 300 400 500 Recoil Energy [keV] (b) Average inelastic energy loss fraction Figure 3-10. Average Frenkel pair production and inelastic energy loss fraction for various cutoff energies 31 3.2. Calibrating MARLOWE to MD In later sections, the results obtained using MARLOWE will be compared against MD calculations. To make meaningful comparisons, we must determine the model parameters in MARLOWE that best represent the physics modeled by the MD simulation. MARLOWE calculates the energy transfer between an incident atom and one in a lattice site via classical mechanics. To dislodge the target atom from its lattice site, the energy transfer must be larger than the binding energy. MARLOWE defines three binding energies for distinct collision types. These are stored in a parameter EBND: EBND(1) is the cohesive energy. If the energies of both the incident and recoil atoms are greater than EBND(1) after the collision, then both particles continue moving and the lattice site is vacated. EBND(2) is the energy for redisplaced atoms. We set this equal to the cohesive energy, EBND(1). EBND(3) is the binding energy for replacement events. In monocrystalline materials where the incident and lattice atoms are the same species, this could be set to zero since the overall state of the lattice is unchanged after the collision. However, since MARLOWE counts each one of these as a displacement, it overestimates the total. EBND(3) can then, in a sense be used as a free parameter to calibrate MARLOWE to MD. Following the work by Vizkelethy and Foiles[18], we calibrate EBND(3) and the re-combination radius in MARLOWE to match the total number of displaced atoms and the number of stable pairs. 3.2.1. Molecular Dynamics Calculations The molecular dynamics calculations are performed using LAMMPS using PKA energies of 2, 5, and 10 keV. The crystalline silicon material was represented as a diamond face-centered cubic lattice structure [19] with a lattice constant of 5.431 angstroms [20].The primitive vectors for the FCC lattice used are a1 = 12 (y + z), a2 = 21 (x + z), and a3 = 12 (x + y). Silicon only has two basis vectors: b1 = (0,0,0), b2 = (0.25,0.25,0.25). The lattice is shown in Figure 3-11. 32 Figure 3-11. Example of crystalline silicon lattice Figure 3-12. ZBL fraction for the silicon Tersoff interatomic potential For silicon, the Tersoff-ZBL potential [21] was used for molecular dynamics modeling. The Tersoff interatomic potential is a simplified many-body potential that has historically been used for displacement cascades. This specific potential was designed for silicon carbide but only the Si-Si interaction of that potential was extracted, similar to other works [18]. Previous work has indicated the need to include short-range repulsive forces within the interatomic potential to account for high-energy collisions [22]. The ZBL potential: 1 𝑍1 𝑍2 𝑒 2 𝑉𝑍 𝐵𝐿 (𝑟) = 𝜙(𝑟/𝑎) (3.2) 4𝜋𝜖 0 𝑟 33 where Z1 and Z2 represent the projectile ion and target atom, r is the radial distance, and phi is the ZBL screening function, was included by blending it in with the interatomic potential using a Fermi distribution as shown in Figure 3-12. A verification of the interatomic atomic potential was done by looking at the cohesive energy as shown in Figure 3-13. The cohesive minimum energy corresponds with the proper lattice constant of crystalline silicon. Figure 3-13. Cohesive Energy of silicon as a function of lattice constant For the BCA-MD calibration, the peak number of displacements was used. This was achieved by keeping track the number of atoms that had displaced more than 1.68 angstroms from their original position. An example of cascade is shown in Figure 3-14. As the PKA ion progresses through the material, subcluster pockets begin to form. It should be stated that the number of displacements does not take into consideration any replacement collisions. The reduction in number of displacements shown in Figure 3-15 comes from the recombination of atoms with their original lattice sites. 34 Figure 3-14. Example of a cascade in silicon Since the peak width for the displacements can be narrow because of short-term recombination, the LAMMPS simulation has to be tallied at every 100 time steps. Looking at Figure 3-15, the peak of the displacement cascade can be formed and annihilated within less than a picosecond. Time steps are determined by the most energetic ion in the system for these simulations. 5 keV 3 keV 250 1 keV 200 Displacements 150 100 50 0 10−1 100 101 Time (ps) Figure 3-15. Number of displacements vs time for various initial PKA events 35 Additionally, the number of vacancies/interstitials were determined using the Wigner-Seitz defect model within OVITO [23]. This methodology looks at comparing the lattice before and after the displacement cascade. Using the pre-irradiated lattice as the reference, the Wigner-Seitz cells are defined. Sites with no atoms are considered vacancies and sites with more than 1 atom contributed to the total interstitial count. Figure 3-16. Example of the Wigner-Seitz method for a 5 keV cascade in silicon 36 3.2.2. Calibrating Displacements The first step is to calibrate MARLOWE’s EBND(3) parameter such that the total number of displaced atoms reproduces the peak number of displacements in the MD simulation. The total number of displaced atoms is simply the number of atoms that had enough energy to leave the lattice site. This quantity is insensitive to recombination effects. The ZBL potential is used in MARLOWE. The other physics options are set to reflect the MD modeling: the crystalline geometry and the non-local inelastic energy loss model. A 1 eV cutoff energy is used. Figure 3-17 plots the total number of displaced atoms for various EBND(3) values as a√︁function of PKA energy. The error bars on the MD results are the standard error of the mean (𝜎/ 𝑁 𝑠𝑎𝑚 𝑝𝑙𝑒𝑠 ). These errors are due to the relatively small number of MD calculations that it was possible to run for this work. The error bars have been omitted from the MARLOWE calculations since it is trivial to perform enough calculations to make the statistical errors on the mean negligible. From this plot, it appears that an EBND(3) value between 2 and 2.5 eV provides a reasonable fit to the MD results. Our intent here is to yield Frenkel pair probability density functions which are comparable to MD to provide a better level of confidence in the BCA results for more complicated quantities later in this report. A detailed fit of the EBND(3) parameter would require many more expensive MD calculations, which is beyond the scope of these aims. Here, we simply take EBND(3) = 2 to be a reasonable approximation. 3.2.3. Calibrating Stable Pairs Since the BCA calculation has no concept of thermal recombination, the effect can be approximated by setting a recombination radius. All displaced atoms that are terminated within a given radius of their lattice sites will be assumed to quickly recombine and are therefore omitted from the stable pair count. The recombination range is varied between 2-7 Å. The previously obtained value of EBND(3) = 2 is used throughout. The recombination range is applied by post-processing the "Separations of Distant Interstitial-Vacancy Pairs" table in MARLOWE, counting only entries with distances greater than the threshold. The average number of stable pairs as a function of PKA energy for various recombination ranges is shown in Figure 3-18. The error bars on the MD calculations are once again the standard error of the mean number of stable pairs at a time after the number of defects has reached equilibrium. 3.2.4. Calibration Summary The MD-calibrated model options and parameters in MARLOWE are summarized in Table 3-5 37 600 1.0 eV 1.5 eV 2.0 eV 2.5 eV 500 MD Total Displaced Atoms 400 300 200 100 2 3 4 5 6 7 8 9 10 Recoil Energy [keV] Figure 3-17. The effect of varying EBND(3) on the total number of displaced atoms predicted by MARLOWE in comparison to MD as a function of PKA energy Using the calibrated options in Table 3-5, the probability density functions for the total number of displacements and the number of stable Frenkel pairs predicted by MARLOWE are shown in Figures 3-19 and 3-20. The MARLOWE distributions are seen to be in good agreement with the MD results obtained using LAMMPS. While it was not feasible to run as many MD calculations, which led to poorer statistical convergence at probability densities below around 0.001; the results still clearly show a similar trend in the tails as predicted by MARLOWE. Due to the drastic increase in computational expense associated with increasing the primary recoil energy, the calibration was performed only for PKAs with energies below 10 keV. While much higher PKA energies will be considered, it is not unreasonable to assume that these calibrations would be applicable to higher energy recoils. It has been suggested[24] that while higher-energy recoils generally produce more cascades, the structure of the individual end-of-track cascades, once an ion’s energy decreases to around 10 keV are identical. Figure 3-21 plots the stable pair production against inelastic energy loss for 35,000 calibrated MARLOWE calculations. 38 200 Recombination Range [Å] 2 3 175 4 4.5 4.75 150 5 6 Stable Pairs 125 7 MD 100 75 50 25 2 3 4 5 6 7 8 9 10 Recoil Energy [keV] Figure 3-18. The effect of varying the recombination range on the total number of Frenkel pairs predicted by MARLOWE in comparison to MD as a function of PKA energy Table 3-5. MD-Calibrated Model Options for Si-Si calculations in MARLOWE Option Setting Notes Interatomic Potential Screened Coulomb Execute MARLOWE with -i expsum Inelastic Energy Loss ZBL Execute MARLOWE with -p zblloss Inelastic Loss Option Non-local INEL=1 in &ATOM section Composition Crystalline POLY=0 in &XTAL section Cutoff Energy 1 eV EQUIT=1 in &ATOM section EBND(3) 2 eV EBND(3)=2 in &ATOM section Recombination Range 4.75 Å Process "Separation Distances" table 39 2 keV 5 keV 10 keV 10−1 10−1 10−1 10−2 10−2 10−2 10−3 10−3 10−3 10−4 10−4 10−4 10−5 10−5 10−5 50 100 150 0 200 0 250 500 Total Displaced Atoms MARLOWE LAMMPS Figure 3-19. Total Displaced Ions - Comparison of probability density functions calculated using calibrated MARLOWE and MD 40 2 keV 5 keV 10 keV 100 100 100 10−1 10−1 10−1 10−2 10−2 10−2 10−3 10−3 10−3 10−4 10−4 10−4 10−5 10−5 10−5 0 20 40 0 50 100 0 100 200 Stable Pairs MARLOWE LAMMPS Figure 3-20. Stable Pairs - Comparison of probability density functions calculated using calibrated MARLOWE and MD 41 Figure 3-21. Correlation between inelastic loss fraction and stable pair formation for various PKA energies. 42 3.3. Distribution Fits The ‘stats‘ module within the Python SciPy package contains a large number of continuous statistical distributions, along with functions to sample from them, fit them to data, and return the probability density function given a set of parameters. All 107 continuous statistical distributions from version 1.11.1 of Scipy.stats were considered. The Maximum Likelihood Estimation (MLE) method was used to fit each distribution to the MARLOWE-calculated results. The goodness-of-fit of the distributions were then ranked by the p-values obtained through the Kolmogorov-Smirnov test. The ten probability density functions with the highest p-values for the calculated Frenkel Pair values are plotted in Figure 3-23. While the Asymmetric Laplace distribution scores highly for both stable pairs and inelastic energy losses, it exhibits an unphysical inflection point about the location parameter. The Generalized Hyperbolic function also scores highly but is discarded as it is a rather complex function requiring five parameters to be fitted. The generalized logistic distribution ("genlogistic") is selected for its ability to describe Frenkel pair and inelastic loss probabilities with a simple 3-parameter form. With the location (𝜇) and scale (𝜎) parameters applied, the generalized logistic distribution takes the form: 𝐹 (𝑥; 𝜇, 𝜎, 𝛼) = (1 + 𝑒 −(𝑥+𝜇) ) −𝛼 /𝜎 (3.3) Figure 3-24 shows the best-fit parameters for the generalized logistic function as a function of recoil energy. The variation in parameter values with recoil energy is observed to be relatively smooth for the stable pair distribution. The location and scale parameters follow an exponential-like trend with increasing recoil energy. The skew parameter 𝛼 shows a decrease for recoil energies up to 25 keV, which then increases and behaves linearly for recoils greater than 400 keV. The parametric fits to the inelastic energy loss fraction appear somewhat more chaotic at first glance, but this is largely due to the normalization inherent in expressing this quantity as a fraction of the primary recoil energy. Without this normalization, the trends of the location and scale parameters are similar to those for the stable pairs. In this case, the behavior of the skew parameter shows the inverse behavior of that for the stable pairs in that it increases briefly up to 25 keV and then steadily declines, becoming a linear decrease for recoils greater than 400 keV. 43 10−1 10−2 10−3 MARLOWE genhyperbolic norminvgauss 10−4 johnsonsu laplace asymmetric genlogistic skewcauchy 10−5 loggamma johnsonsb gumbel l rel breitwigner 10−6 0 20 40 60 80 100 120 Stable Pairs (a) 5 keV 10−2 10−3 10−4 MARLOWE 10−5 laplace asymmetric genhyperbolic genlogistic 10−6 johnsonsu gumbel l loggamma logistic 10−7 johnsonsb gennorm dgamma 10−8 0 2000 4000 6000 8000 Stable Pairs (b) 1 MeV Figure 3-22. The ten best-fit probability density functions for the number of stable Frenkel Pairs for two PKA energies 44 102 MARLOWE norminvgauss genhyperbolic 101 burr12 johnsonsu mielke fisk laplace asymmetric 100 burr powerlognorm loglaplace 10−1 10−2 10−3 0.0 0.2 0.4 0.6 0.8 1.0 (a) 5 keV 102 MARLOWE laplace asymmetric genhyperbolic 101 burr12 johnsonsu norminvgauss mielke burr 100 genlogistic fisk loglaplace 10−1 10−2 10−3 0.5 0.6 0.7 0.8 0.9 1.0 (b) 1 MeV Figure 3-23. The ten best-fit probability density functions for the inelastic energy loss fraction for two PKA energies 45 Stable Pairs Inelastic Energy Loss 104 107 106 103 105 Parameter Value 102 104 101 103 102 100 101 10−1 100 10−1 101 103 10−1 101 103 Recoil Energy [keV] Recoil Energy [keV] µ σ α Figure 3-24. Fit parameters of the generalized logistic function for a range of recoil energies 46 3.3.1. Alternate Fits Alternate methods were also explored for determining the probability density functions for the MARLOWE-calculated quantities. These methods are described in the following sections. 3.3.1.1. Kernel Density Estimator The parametric distribution functions investigated in the previous section were shown to produce good fits around the peaks of the distribution, but none were able to fit the multi-modal behavior exhibited for low numbers of stable pairs, or high inelastic energy loss fractions. A Gaussian Kernel Density Estimator (KDE) was applied to the MARLOWE results in an attempt to improve upon the fits obtained earlier. Figure 3-25 shows the KDE fits to the MARLOWE results for several recoil energies. 10−1 101 10−2 100 10−3 10−1 10−4 10−5 10−2 100 101 102 103 104 0.5 0.6 0.7 0.8 0.9 1.0 Stable Pairs Inelastic Energy Loss Fraction 5 keV 500 keV 50 keV 1 MeV Figure 3-25. Kernel Density Estimator fits (smooth) to MARLOWE results (histograms) Scott’s rule is used to determine the bandwidth of the Gaussian kernels which appears to have produced reasonable probability density functions for the MARLOWE results, with however a little over-fitting observed in the inelastic energy loss probabilities. While producing favorable fits to the MARLOWE results, a drawback to this non-parametric approach include a lack of clear method to describe trends across different recoil energies. The method is also somewhat sensitive to the choice of bandwidth. 47 3.3.1.2. TableCurve2D TableCurve 2D[25] was also investigated as a discovery tool for the optimal analytic distribution. Instead of fitting directly to the calculation results, this technique relies on first binning the results and then converting them to a density function by dividing by the bin widths. The technique is then a simple regression of the trial function parameters to the calculated probability density function. Similarly to the Python approach, TableCurve 2D also suggested a logistic distribution when supplied with the calculated probability density function for the number of stable pairs (see Figure 3-26). While providing a useful verification of the SciPy.stats technique, it was less convenient to integrate into a scripted process and was therefore not used to calculate the parameters over the large range of PKA energies considered here. Figure 3-26. TableCurve 2D fit to stable pair distribution 3.4. Recoil Distributions The probability distributions for Frenkel pairs and inelastic energy losses for incident neutrons will be explored in this section. The recoil distributions used for 1 and 14 MeV neutrons are shown in Figure 3-27. These were calculated using EMPIRE and include not only Si28 recoils but also other heavy ion species of similar mass produced through transmutation reactions: Si29 , Al27 , Al28 , Mg24 , and Mg25 . In MARLOWE, these are all modeled as Si28 , which is assumed to be representative of all heavy ion recoils due to its relative abundance and similar mass to the other species. 48 1 MeV Neutron 14 MeV Neutron 10−2 10−4 10−6 Probablity 10−8 10−10 10−12 10−13 10−11 10−9 10−7 10−5 10−3 10−1 101 Ion Energy [MeV] Figure 3-27. Si recoil distributions for 1 and 14 MeV incident neutrons 3.4.1. Sampling Method Codes such as CASCADE sample ion energies directly from the recoil distribution of an incident neutron spectrum. These energies are then used to initialize MARLOWE calculations to calculate the average Frenkel pair production for a given neutron environment. Since this method samples ion energies directly from their probability distribution, the statistics quickly converge around the most probable ion energies. Another method is to sample ion energies uniformly to produce a database of MARLOWE cal- culations. Since these samples are independent of any recoil distribution, they can be binned and weighted to the recoil probabilities of practically any incident neutron energy (provided the database covers the extents of the recoil energies). Figure 3-28 plots the recoil distribution for 14 MeV neutrons and the histograms produced using the two methods and their associated relative errors. Ten thousand samples are drawn for both the rejection sampling technique and the weighted loguniform distribution method. The rejection sampling technique is seen to produce a slightly better agreement with the expected distribution around the peak; however, it did not draw a single sample from the recoil energy bins above 3.6 MeV. At the lower energy side, the rejection sampling technique is shown to be progressively noisier and then fails to draw any samples below 10 eV. While the rejection sampling technique provides good convergence on the mean and standard deviation of the quantities of interest for a given recoil distribution, it is less suited to sampling rare 49 100 10−2 Probability 10−4 10−6 Rejection Sampling from Recoil Distribution Weighted Samples from LogUniform Distribution Expected distribution 10−8 10−6 10−5 10−4 10−3 10−2 10−1 100 101 Recoil Energy [MeV] Figure 3-28. Rejection Sampling vs. Weighted Uniform Samples and possibly consequential events - for example, recoil ions in the high-energy tail of the 14 MeV distribution. Since these events are of interest for this work, the uniform weighting method is used for all subsequent incident neutron calculations. 3.4.2. Probability Density Functions for Incident Neutrons Using the sampling method described above, 300,000 PKA events are simulated in MARLOWE. The sampled PKA energies are then binned into the same energy group structure as the recoil spectrum, and their contributions to the total quantity of interest (number of pairs or inelastic energy loss) are weighted accordingly. In addition to the MARLOWE samples from the recoil distribution, the probability density functions are estimated using the weighted integral of the generalized logistic distributions for each ion energy: ∫ ∞ 𝑃𝐷𝐹 (𝑥) = 𝑓 (𝑥; 𝜇(𝐸), 𝜎(𝐸), 𝛼(𝐸))𝑅(𝐸)𝑑𝐸 (3.4) 0 where 𝜇(𝐸), 𝜎(𝐸), and 𝛼(𝐸) are the fit parameters for the distributions at each ion energy 𝐸, 𝑓 () is the probability density function for the generalized logistic distribution, and 𝑅(𝐸) is the recoil distribution for an incident neutron energy. Equation 3.4 is integrated numerically by creating an energy grid of 5,000 logarithmically spaced points and using log-log interpolation on the parameters and the recoil distributions. 50 Figures 3-29 and 3-30 plot the probability density functions for Frenkel pair numbers and inelastic energy loss for incident neutrons of 1 and 14 MeV, respectively. Frenkel Pairs Inelastic Energy Loss 10−2 102 MARLOWE Functional Form Probability 10−3 101 10−4 100 100 102 104 10−6 10−4 10−2 100 # Pairs Inelastic Energy Loss [MeV] Figure 3-29. Damage quantities for 1-MeV incident Neutrons For inelastic energy loss, except for some minor oscillations at low energies, the functional form is able to predict the MARLOWE probability density distribution for both neutron energies. This is likely due to the individual ion distributions being well-fit by the generalized logistic distribution around the peak, while the features that are less well-fit are less important. The functional form does a reasonable job of describing the probability density function for Frenkel pairs for the 1 MeV incident neutron. The probability density function for Frenkel pairs becomes increasingly bi-modal at higher ion energies (see Figure 3-23) which is not particularly well fit by the analytic PDFs. This leads to a failure to predict the "shoulder" feature in the 1 MeV results, which is more stark in the 14 MeV results. More importantly, however, for low-probability, high-consequence events, excellent agreement is shown for both neutron energies between the functional form and the MARLOWE results at the high-pairs and high-inelastic energy loss end of the distributions. 3.4.2.1. SRIM Comparisons Finally, we compare the MARLOWE probability density functions functions against SRIM. 20,000 silicon ions are sampled from the 14 MeV neutron recoil distribution to initialize the SRIM calculations. Two methods were employed to evaluate Frenkel pair generation: 1) the extraction of 51 Frenkel Pairs Inelastic Energy Loss 10−3 MARLOWE Functional Form 101 Probability 10−4 100 10−5 10−1 100 102 104 10−6 10−4 10−2 100 # Pairs Inelastic Energy Loss [MeV] Figure 3-30. Damage quantities for 14-MeV incident Neutrons defect distributions from the raw VACANCY.txt output in full-cascade mode, and 2) the method suggested by Stoller [4], in which the inferred damage energy of each ion simulation is converted into a defect number via the NRT displacement model formula. As Stoller’s method pertains only to atomic displacements, this method is omitted from comparisons of inelastic energy loss, which is reported directly by SRIM. The SRIM results plot against MARLOWE and the analytic form are shown in Figure 3-31. For inelastic energy loss, excellent agreement is shown between SRIM, MARLOWE, and the functional form. The divergence of SRIM at lower energy losses is primarily due to statistical noise. For the number of pairs, the agreement is somewhat worse. The default SRIM displacement treatment predicts a somewhat higher number of pairs on the right side of the distribution. Using the Stoller method to determine the number of vacancies brings the high-pair peak into alignment with the MARLOWE results, but the amplitude remains higher. 52 Frenkel Pairs Inelastic Energy Loss 10−3 101 Probability 10−4 100 MARLOWE Functional Form SRIM SRIM [Stoller] 10−5 10−1 100 102 104 10−6 10−4 10−2 100 # Pairs Inelastic Energy Loss [MeV] Figure 3-31. Probability Density Functions for 14 MeV incident Neutrons, comparisons with SRIM 53 54 4. SPATIAL DISTRIBUTION In the previous section, the intensity distribution for stable defects and ionization in bulk silicon was studied using the BCA codes SRIM and MARLOWE. The sensitivity of the distribution has been studied for the model options: crystalline vs. amorphous, and local vs. non-local inelastic loss model, and the energy cut off. High fidelity molecular dynamics (MD) calculations were run in bulk silicon, and the results used to tune the binding energies and recombination in radii to calibrate the faster running binary collision approximation (BCA) code MARLOWE. The intensity distributions for the total number of displacements and the number of stable Frenkel pairs were compared between MD and MARLOWE simulations to verify the effect of channeling in a crystalline material. A generalized distribution, described by the generalized logistic distribution, was found to be a reasonable fit to the calculated probability density functions. In modern, small feature-size devices such the gate-all-around nanosheet (GAA NS) FET, containing individual layers as thin as 5 nm (about 15 silicon atoms) [26], the characterization of damage in bulk silicon becomes less relevant. A single recoil event, even of modest energy, can readily pass through all active channels of a device that is miniaturized to such an extent, as can be verified through simple ion range estimations in SRIM. To address this class of device, we apply a probabilistic, parametric approach to assess the likelihood of defect concentrations at various threshold levels being introduced simultaneously to multiple channels of the GAA NSFET device from a single neutron collision. Given the extreme thinness of the channels, it is postulated that the defect concentration can be translated into a degree of disruption of the operational viability of a channel, which at some point ceases to function properly. Other mechanisms of degradation are also possible under irradiation, although GAA NSFETs have been found to be highly resilient to total ionizing dose effects [26]. 4.1. FinFET Geometry In this work, we use a finFET as a representative example of a small feature size device that could be adversely affected by single-event displacement damage. Specifically we consider a Gate-All-Around Nanosheet (GAA NS) device[27]. The dimensions of the sensitive region of the GAA NS device are inferred from schematics[28] and micrograph images[29] in the literature. The dimensions of the model are shown in Figure 4-1. The channel layers are all assumed to be 5 nm in depth, as are each of the intervening layers of gate metal. The length of the sheets is estimated at 50 nm from the micrograph images, and their width is estimated at 10 nm. 55 Figure 4-1. Longitudinal and cross sectional dimensions of the GAAFET NS model. The highlighted region is modeled in MARLOWE As for the materials, the channels are reported in the literature are typically Si for nFETs and SiGE for pFETs. For simplicity of modeling and for consistency with the earlier sections of this report, we set the channel material to Si. The gate metal is typically proprietary, but metals such as aluminum[30], titanium, tantalum, and gold[31] have been reported in the literature. For this work we limit the scope to Aluminium as potentially the worst case in terms of having the lowest stopping powers for Si ions. 4.2. Modeling The multi-layer input in MARLOWE is somewhat complex and is included in Appendix A. The Si channel in which the PKA is generated is selected randomly. Once a channel has been selected, a location within that channel is sampled randomly from the volume. The launch angle is also sampled uniformly via the cosine weighting method. A slab geometry is used in MARLOWE with finite lateral dimensions to represent the material stack-up shown in Figure 4-1. The physics parameters are set according to the calibration performed in the previous section (Table 3-5). Some example cascades for various PKA energies are shown in Figure 4-2. These plots are flattened along the length dimension to enable plotting in 2D. A 3D example for the 5 keV cascade is shown in Figure 4-3. 56 2 keV 5 keV 30 Si 30 Al Z [nm] 20 20 10 10 0 0 4 2 0 2 4 4 2 0 2 4 50 keV 100 keV 30 30 20 20 Z [nm] 10 10 0 0 4 2 0 2 4 4 2 0 2 4 Y [nm] Y [nm] Figure 4-2. Example Cascades for various PKA energies. The plots show the final locations of the interstitial atoms of each species. 4.2.1. Multi-channel Defects We define the probability density function for multi-channel defects as a function of the threshold value for the number of defects that could be tolerable in a given layer. Given a threshold value for defects, we define an indicator variable to signal if the number of defects in layer 𝑖, 𝐷 𝑖 , is equal to or greater than the threshold 𝑇: ( 1 if 𝐷 𝑖 ≥ 𝑇 𝐼𝑖 = 0 if 𝐷 𝑖 < 𝑇 the probability of two or more defects defects can then be written: 𝑃(𝐼1 + 𝐼2 + 𝐼3 ≥ 2). The probability density function for seeing two or more layers simultaneously meeting or exceeding a threshold (𝑥) number of defects can be written: 𝑓2 (𝑥) = 𝑃(𝐼1 + 𝐼2 + 𝐼3 ≥ 2|𝑇 = 𝑥) and similarly for all three layers: 57 Si Al 35 30 25 20 Z [nm] 15 10 5 0 20 4 10 2 X [nm0 0 ] ] 10 20 4 2 Y [nm Figure 4-3. 3D plot of the 5 keV cascade shown in Figure 4-2 𝑓3 (𝑥) = 𝑃(𝐼1 + 𝐼2 + 𝐼3 = 3|𝑇 = 𝑥) The probability density functions for 𝑓2 and 𝑓3 are calculated using MARLOWE, with 20,000 ions launched for each PKA energy. The probabilities for ≥ 2 and 3 simultaneously-affected layers are plotted in Figure 4-4. From these results, it can be seen that for a low number of stable pairs, the probability densities are somewhat similar across all the ion energies. The probability of a single defect in two or more layers is around 0.1 ± around 50% for all the ion energies considered here. Similarly, the probability for a single defect in three layers is around 0.01 ± around 60% for all ion energies. Starker differences are observed in the tails of the distributions. For the 2 keV case, no more than 13 defects in two or more layers were observed simultaneously in the 20,000 ion tracks sampled. For the 50 keV case, at least 130 defects were observed in two layers simultaneously, and 53 in three layers simultaneously. From these PKA energies, 50 keV appears to be the worst-case for simultaneous defect production in multiple layers. At lower energies, the PKAs lack the energy to initialize multiple cascades in multiple layers simultaneously. At higher energies, the ions have longer ranges, and are more likely to lose a greater fraction of their energy through inelastic processes rather than elastic collisions over the size of the GAAFET (for PKAs produced within the volume), so the probability of initializing cascades in each of the regions is reduced. At 50 keV, it appears a balance has been reached between 58 2 keV 5 keV Probablity ≥ Num. Defects in N layers simultaneously ≥ 2 Layers 10−2 = 3 Layers 10−2 10−5 10−5 50 100 150 50 100 150 10 keV 50 keV 10−2 10−2 10−5 10−5 50 100 150 50 100 150 100 keV 500 keV 10−2 10−2 10−5 10−5 50 100 150 50 100 150 Num. Defects Num. Defects Figure 4-4. Probability density functions for stable pair formation in multiple layers for a range of PKA energies the energy available to seed cascades and the energy at which elastic collisions are more likely to occur. 4.2.2. Comparisons with SRIM Similar SRIM calculations are performed to verify the results produced by MARLOWE. Due to the way SRIM defines the geometry, it was simpler in this case to assume that all ions were produced in the first silicon layer of the geometry. It is also not possible to define the lateral extents in SRIM in the same way as can be done in MARLOWE, so a pure slab geometry is assumed with infinite lateral extents. Figure 4-5 plots the comparisons with SRIM, for the nominal MARLOWE case, and MARLOWE calculations that have been "SRIMized" to be more aligned with the SRIM model. This means the lateral extents are set to be essentially infinite, and all PKAs are sampled from the first layer of silicon. The crystal model in MARLOWE is retained, however, in contrast to the amorphous geometry assumed in SRIM. It can be seen that, in general, SRIM predicts significantly higher probabilities of multi-layer defects. At higher recoil energies, the SRIM calculations continue to predict higher probabilities 59 2 keV 5 keV 10−2 10−2 Probablity ≥ Num. Defects in N layers simultaneously 10−5 10−5 20 40 60 80 100 20 40 60 80 100 10 keV 50 keV 10−2 10−2 10−5 10−5 100 200 300 400 500 100 200 300 400 500 100 keV 500 keV 10−2 10−2 10−5 10−5 200 400 600 800 1000 200 400 600 800 1000 Num. Defects Num. Defects N ≥ 2 layers N = 3 layers Nominal MARLOWE SRIMized MARLOWE SRIM Figure 4-5. Probability density functions for stable pair formation in multiple layers for a range of PKA energies, comparison between MARLOWE, "SRIMized" MARLOWE, and SRIM of numerous multilayer defects whereas the MARLOWE results peak around 50 keV. A primary driver of this difference is simply the usage of an infinite slab geometry in SRIM. In the nominal MARLOWE calculations, the PKAs launched isotropically in angle have a good chance of exiting one of the lateral dimensions before crossing into another silicon layer, while in an infinite slab, all PKAs with an initial velocity vector pointing towards the other layers must be stopped or deflected to prevent a multi-layer effect. For a fairer comparison, the "SRIMized" MARLOWE calculations implement infinite lateral dimen- sions. At low PKA energies, the lateral dimensions do not appear to play a role in the probabilities for multi-layer defects. This is an intuitive result when considering the ranges of the lower energy ions. Between 10 and 100 keV, the "SRIMized" MARLOWE results are in good agreement with the SRIM results for low ( 100) numbers of simultaneous defects. At higher defect numbers the results diverge, with SRIM predicting much higher probabilities. At smaller defect thresholds, the probability of seeing multiple-layer effects is likely more dependent 60 on the cascade dynamics and stopping powers in the silicon and aluminum layers. For high defect thresholds, the probabilities are more sensitive to the overall production intensity. To verify this statement, we vary the 𝐸 𝑑 parameter in the SRIM results such that the overall stable pair intensity in bulk silicon is in agreement with the MD-calibrated MARLOWE results shown in Figure 3-18. The default 𝐸 𝑑 for silicon is 15 eV. For a 10 keV ion, SRIM predicts on average 240 stable pairs per recoil in bulk material. The MD-calibrated MARLOWE prediction for stable pair production is around 150 per recoil. Modifying 𝐸 𝑑 in SRIM to 25 eV reduces the number of stable pairs per recoil to 145, which is in line with the MARLOWE results. 100 Probablity ≥ Num. Defects in N layers simultaneously 10−1 10−2 10−3 10−4 10−5 20 40 60 80 100 120 140 Num. Defects N ≥ 2 layers N = 3 layers MARLOWE SRIM Ed = 15 eV SRIM Ed = 25 eV Figure 4-6. Probability density functions for stable pair formation in multiple layers for 10 keV recoil ions Figure 4-6 compares SRIM calculations using threshold displacement energies of 15 and 25 keV and the "SRIMized" MARLOWE results. The results support the previously stated hypothesis, as the agreement at low defect thresholds is not significantly changed by increasing the threshold displacement energy. However, the agreement at higher defect thresholds is improved. 61 4.2.3. Recoil Distributions A database of 300,000 MARLOWE calculations is generated using random ion energies sampled via the method described in Section 3.4.1. Each MARLOWE calculation is then weighted via log-log interpolation of the incident neutron recoil spectrum, which is used in the calculation of 𝑓2 (𝑥) and 𝑓3 (𝑥). For example 𝑓2 (𝑥) is calculated from the MARLOWE results as: ( Í𝑁𝑐𝑎𝑙𝑐𝑠 𝑤 𝑐 if 𝐼 (𝑥) ≥ 2 𝑐 0 if 𝐼 (𝑥) < 2 𝑓2 (𝑥) = Í𝑁𝑐𝑎𝑙𝑐𝑠 𝑐 𝑤𝑐 where 𝑤 𝑐 is the log-log interpolated weight of a given MARLOWE calculation based on the PKA energy and the recoil spectrum for an incident neutron, and I(x) is a function indicating how many of the three layers meet or exceed the threshold (𝑥) defect count: 3 ( ∑︁ 1 if 𝐷 𝑖 ≥ 𝑥 𝐼 (𝑥) = 𝑖 0 if 𝐷 𝑖 < 𝑥 Since loguniform sampling was used to sample recoil energies uniformly across the logarithmic domain, and additional factor must be applied to weight each sample to reverse the imposed bias towards lower energies. The loguniform samples are weighted as: 1 𝑤𝑖 = 𝑥𝑖 (𝑙𝑜𝑔(𝑏) − 𝑙𝑜𝑔(𝑎)) where 𝑥𝑖 is the energy of the random loguniform sample. Figure 4-7 plots the probability density functions for multiple simultaneous defects in 2 and 3 layers simultaneously for incident neutrons of 1 and 14 MeV. It can be seen that the incident 1 MeV neutrons have generally higher probabilities of producing damage in multiple layers than 14 MeV neutrons, which is consistent when considering the recoil distributions and the results for single-ion energies shown in Figure 4-4. Here the PKA location, energy and launch angle are sampled independently. While these results are reasonable for an isotropic neutron environment, for planar environments, one could contrive a realistic worst-case by correlating the recoil energies and the launch angles. For example, one could assume that the planar neutron angle of incidence is perpendicular to the first silicon layer in the geometry. For elastic reactions, it can be reasonably assumed that the launch angle of the PKA is strongly correlated with the neutron angle of incidence, while ions produced by inelastic processes are emitted isotropically. Such assumptions increase the probability of recoil ions being emitted in directions that produce damage in all three channels of the finFET. 62 100 N ≥ 2 layers N = 3 layers 1 MeV Neutron Probablity ≥ Num. Defects in N layers simultaneously 14 MeV Neutron 10−1 10−2 10−3 10−4 10−5 10−6 20 40 60 80 100 120 140 160 Num. Defects Figure 4-7. Probability density functions for stable pair formation in multiple layers for 1 and 14 MeV incident neutrons 63 64 5. SUMMARY For neutron-induced total dose effects in semiconducting devices, it is a reasonable approximation to simply take the mean defect production or ionization rates per PKA energy and scale by the incident neutron fluence and spectrum-dependent recoil probabilities. In small-feature size devices, the probability of a neutron event occurring in the sensitive volume decreases, while the concentration of defects produced by a single neutron event within the volume increases. For such devices, the probability densities for a single neutron event can be more important for understanding failure probabilities than the aggregated statistics, which assume numerous neutron events within a sensitive volume. In this report, we document a statistical approach to intensity distributions for neutron-induced defect and ionization production in silicon devices. The BCA technique employed by codes such as MARLOWE and SRIM is approximate, but it is a practical and computationally tractable method for modeling large numbers of recoil cascades, as opposed to expensive MD simulations. Here, we demonstrate the sensitivity of the ionization and displacement cascade to the physics options in MARLOWE. It was shown that not only does the choice of physics model option change the mean ionization and defect quantities, but the shapes of the distributions also change. For example, the amorphous composition option predicted Gaussian-like probability density functions, while the effect of channeling enabled by the crystalline composition produced a non-Gaussian tail in both the defect and ionization probabilities. However, when the crystalline composition model was used in conjunction with the non-local inelastic energy loss model, it was shown that the MARLOWE Frenkel pair production rates could be calibrated to reproduce high-fidelity MD results by tuning only two parameters. Not only did MARLOWE agree with MD on the mean values for the total number of displaced atoms and stable Frenkel pairs, but the probability density functions were also shown to be of a similar shape. The intensity distributions for Frenkel pairs and ionization were found to be reasonably well described by a generalized logistic distribution. Fitting the three parameters of this distribution to MARLOWE probability density functions yielded smooth variations in the parameter values with PKA energy. Interpolating and weighting the generalized logistic distributions for each ion energy according to a neutron recoil distribution produced an analytic function which that in excellent agreement with direct MARLOWE calculations of the probabilities of high-impact displacement and ionization events. With MARLOWE calibrated to yield stable Frenkel pair probabilities in agreement with MD simulations, attention was turned to an example small feature size device. The probability of a single neutron strike in any channel of a GAA NSFET producing defects in multiple channels was calculated for a range of defect thresholds. These probabilities showed an exponential-like decay as the defect threshold was increased. The probability of seeing critical damage in two silicon regions from a single neutron strike varied as a function of the number of defects required to reach 65 the critical threshold, but for low defect numbers, the probability was observed to be around once per 10 neutron events for both 1 MeV and 14 MeV incident energies. For critical damage in three regions from a single neutron event, the probability (for low thresholds) was shown to be around once per 100 neutron events. 5.1. Further Work Building upon the work presented here, a simple extension would be to convolve the neutron interaction cross sections and derive the probabilities presented here for a given neutron fluence and spectrum. Additionally, the feature size in an example device could be parameterized to determine the threshold at which single-event or total dose displacement effects become more important in terms of the probabilities of defect concentrations in multiple sensitive volumes. Additionally, there are a number of threads left to pull in calculating bulk silicon intensity distri- butions with BCA codes. The crystalline structure combined with the non-local inelastic energy loss model was determined to be being most similar to the model used in the MD simulations. The MARLOWE-calculated probability density functions for Frenkel pair production and ionization likely overestimate the long tails of the distributions for two reasons. Firstly, all the lattice ions were considered stationary in the MARLOWE calculations. In reality, these ions will undergo ther- mal vibrations. Secondly, the bulk silicon simulated here was a single, infinite lattice. In reality the silicon may exhibit polycrystalline features due to finite grain sizes in the bulk silicon. Both of these factors limit the maximum range an ion could channel and therefore reduce the probabilities at the low extreme Frenkel pair production rates and the high extreme inelastic energy loss fractions. With regard to molecular dynamics, one of the main concerns for running displacement cascades in LAMMPS was the computational cost. The solution used in this work was to calibrate the MARLOWE parameters using the molecular dynamics results. However, there are higher-fidelity solutions that don’t involve introducing potentially nonphysical calibration parameters. One such solution would be to implement a BCA-MD hybrid methodology. Rather than calibrating the MARLOWE simulation, ions that fall below a certain energy threshold terminate their trajectory within MARLOWE and continue, either through a forward model in LAMMPS or through a library of pre-determined molecular dynamics results. This methodology has been used in the past with metallic materials for structural properties [32] but applying it to nSEE would be a novel approach. Furthermore, if one could improve the energy coupling between BCA and MD, localized temperature increases from ionizing energy loss could be investigated. Specifically, examining if the ionizing energy loss from a high LET silicon ion impacts the molecular dynamics simulation if the electron temperature component from it’s ion track is included. Usually the energy would be decoupled in this workflow, assuming that the range of the ion would make the localized temperature increase negligible for the defect production region, but this has yet to be shown computationally. Going beyond material studies, the feature sizes of state-of-the-art devices are approaching di- mensions in which they can fully be modeled in molecular dynamics. Traditionally, displacement cascades in molecular dynamics often focus on one material. This simplifies the simulation since the interatomic potential often only includes interactions for that material. However, if there are multiple layers of different materials, addressing the interatomic potential becomes complicated. 66 Assuming that you could develop a methodology for keeping the interfaces stable, only the repulsive potential would be necessary to account for atoms leaving their original layer. This is because we can assume that the number of atoms displaced from one layer to the next is much smaller than the total number of atoms in that layer. However, if the layers are not uniform and you introduced layers with large width gradients, then this approximation would break down, and you would need to accurately implement the interatomic potential for all possible material interactions. With the advent of machine-learning-derived interatomic potentials[33], this may be feasible in the near future. 67 68 REFERENCES [1] James F. Ziegler, M. D. Ziegler, and J. P. Biersack. SRIM - The Stopping and Range of Ions in Matter (2010). Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms, 268(11):1818–1823, 2010. [2] M. T. Robinson. The Binary Collision Approximation: Background and Introduction. Radi- ation Effects and Defects in Solids, (1):3–20, 1994. [3] A. P. Thompson, H. M. Aktulga, R. Berger, D. S. Bolintineanu, W. M. Brown, P. S. Crozier, P. J. in ’t Veld, A. Kohlmeyer, S. G. Moore, T. D. Nguyen, R. Shan, M. J. Stevens, J. Tranchida, C. Trott, and S. J. Plimpton. LAMMPS - A Flexible Simulation Tool for Particle-Based Materials Modeling at the Atomic, Meso, and Continuum Scales. Computer Physics Com- munications, 271:108171, 2022. [4] R. E. Stoller, M. B. Toloczko, G. S. Was, A. G. Certain, S. Dwaraknath, and F. A. Garner. On the Use of SRIM for Computing Radiation Damage Exposure. Nuclear Instruments and Methods in Physics Research B, 310:75–80, 2013. [5] W. J. Weber and Y. Zhang. Predicting Damage Production in Monoatomic and Multi- Elemental Targets Using Stopping and Range of Ions in Matter Code: Challenges and Recommendations. Current Opinion in Solid State and Materials Science, 23:100757, 2019. [6] G. Demange, E. Antoshchenkova, M. Hayoun, L. Lunéville, and D. Simeone. A Rational Use of BCA Code MARLOWE for Ballistic Effects of Ion Beam Irradiation in the Ion Mixing Formalism: Comparison to Molecular Dynamics, 2016. [7] C. J. Ortiz, A. Souidi, C. S. Becquart, C. Domain, and M. Hou. Recent Radiation Damage Studies and Developments of the MARLOWE Code. Radiation Effects and Defects in Solids, 169(7):592–602, 2014. [8] M. Hou, C. J. Ortiz, C. S. Becquart, C. Domain, U. Sarkar, and A. Debacker. Microstructure Evolution of Irradiated Tungsten: Crystal Effects in He and H Implantation as Modeled in the Binary Collision Approximation. Journal of Nuclear Materials, 403(1):89–100, 2010. [9] David K. Brice. Ziegler-Biersack-Littmark Inelastic Energy Loss Implemented in the MAR- LOWE Code. Internal Memo. [10] Dávid Lucsányi, Rubén García Alía, Kacper Biłko, Matteo Cecchetto, Salvatore Fiore, and Elisa Pirovano. G4SEE: A Geant4-Based Single Event Effect Simulation Toolkit and Its Validation Through Monoenergetic Neutron Measurements. IEEE Transactions on Nuclear Science, 69(3):273–281, 2022. 69 [11] Kevin Warren, Robert Reed, Robert Weller, Marcus Mendenhall, Brian Sierawski, and Ronald Schrimpf. Applications of Monte Carlo Radiation Transport Simulation Techniques for Pre- dicting Single Event Effects in Microelectronics. AIP Conference Proceedings, 1336(1):643– 648, 06 2011. [12] Danielius Lingis, Mindaugas Gaspariūnas, Artūras Plukis, and Vidmantas Remeikis. Im- provements and Validation of Particle Channeling Model in GEANT4. Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms, 525:1–12, 2022. [13] Gilles Demange, Ekaterina Antoshchenkova, M. Hayoun, Laurence Lunéville, and D. Sime- one. A Rational Use of BCA Code MARLOWE for Ballistic Effects of Ion Beam Irradiation in the Ion Mixing Formalism: Comparison to Molecular Dynamics. 10 2016. [14] George P. Mueller. The Inadequacies of Damage Energy as a Measure of Displacement Damage. Technical Report NRL Memorandum Report 6443, Naval Research Laboratory, April 1989. [15] G. Demange, E. Antoshchenkova, M. Hayoun, L. Lunéville, and D. Simeone. Simulating the Ballistic Effects of Ion Irradiation in the Binary Collision Approximation: A First Step Toward the Ion Mixing Framework. Journal of Nuclear Materials, 486:26–33, 2017. [16] Seiki Saito et al. Grain Size Dependence of Penetration Depth of Hydrogen Injection into Polycrystalline Graphite by Molecular Simulation. Japanese Journal of Applied Physics, 53, 2014. [17] Fernando Mota, Christophe Ortiz, and Rafael Vila. Tools and Procedures for Radiation Dam- age Modelling and Intercomparison of Experiments. IAEA CRP F44003 Primary Radiation Damage Cross Section, October 2017. [18] Gyorgy Vizkelethy and Stephen M. Foiles. Determination of Recombination Radius in Si for Binary Collision Approximation Codes. Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms, 371:111–115, 2016. The 22nd International Conference on Ion Beam Analysis (IBA 2015). [19] Charles Kittel and R. W. Hellwarth. Introduction to Solid State Physics. Physics Today, 10(6):43–44, 1957. [20] P. Becker, P. Scyfried, and H. Siegert. The Lattice Parameter of Highly Pure Silicon Single Crystals. Zeitschrift für Physik B Condensed Matter, 48:17–21, 1982. [21] R. Devanathan, T. Diaz de la Rubia, and W. J. Weber. Displacement Threshold Energies in B-SiC. Journal of Nuclear Materials, 253(1):47–52, 1998. [22] R. E. Stoller, A. Tamm, L. K. Béland, G. D. Samolyuk, G. M. Stocks, A. Caro, L. V. Slipchenko, Yu. N. Osetsky, A. Aabloo, Mattias Klintenberg, et al. Impact of Short-Range Forces on Defect Production from High-Energy Collisions. Journal of Chemical Theory and Computation, 12(6):2871–2879, 2016. 70 [23] A. Stukowski. Visualization and Analysis of Atomistic Simulation Data with OVITO–The Open Visualization Tool. Modelling and Simulation in Materials Science and Engineering, 18(1), Dec. 2009. Art. no. 015012. [24] Christopher Race. A Radiation Damage Cascade, pages 9–13. Springer Berlin Heidelberg, Berlin, Heidelberg, 2011. [25] Download TableCurve2D Software. https://systatsoftware.com/ download-tablecurve2d-software/. [Accessed 23-07-2025]. [26] M. Titze, A. Belianinov, A. Tonigan, S. Su, G. Vizkelethy, W. Wampler, B. Hehr, M. Wang, H. Zhou, V. Narayanan, E. S. Bielejec, and R. Arghavani. Displacement Damage, Total Ionizing Dose, and Transient Ionization Effects in Gate-All-Around Field Effect Transistors. ACS: Applied Electronic Materials, 6:5759–5765, 2024. [27] N. Loubet, S. Kal, C. Alix, S. Pancharatnam, H. Zhou, C. Durfee, M. Belyansky, N. Haller, K. Watanabe, T. Devarajan, J. Zhang, X. Miao, M. Sankar, M. Breton, R. Chao, A. Greene, L. Yu, J. Frougier, D. Chanemougame, K. Tapily, J. Smith, V. Basker, A. Mosden, P. Biolsi, T. Q. Hurd, R. Divakaruni, B. Haran, and H. Bu. A Novel Dry Selective Etch of SiGe for the Enablement of High Performance Logic Stacked Gate-All-Around NanoSheet Devices. In 2019 IEEE International Electron Devices Meeting (IEDM), pages 11.4.1–11.4.4, 2019. [28] Huimei Zhou, Miaomiao Wang, Ruqiang Bao, Curtis Durfee, Liqiao Qin, and Jingyun Zhang. SiGe Gate-All-Around Nanosheet Reliability. In 2022 IEEE International Reliability Physics Symposium (IRPS), pages P60–1–P60–4, 2022. [29] S. Mochizuki, B. Colombeau, J. Zhang, S. C. Kung, M. Stolfi, H. Zhou, M. Breton, K. Watan- abe, J. Li, H. Jagannathan, M. Cogorno, T. Mandrekar, P. Chen, N. Loubet, S. Natarajan, and B. Haran. Structural and Electrical Demonstration of SiGe Cladded Channel for PMOS Stacked Nanosheet Gate-All-Around Devices. In 2020 IEEE Symposium on VLSI Technology, pages 1–2, 2020. [30] Yujia Zhai, Leo Mathew, Rajesh Rao, Marylene Palard, Sonali Chopra, John G. Ekerdt, Leonard F. Register, and Sanjay K. Banerjee. High-Performance Vertical Gate-All-Around Silicon Nanowire FET with High- 𝜅 /Metal Gate. IEEE Transactions on Electron Devices, 61(11):3896–3900, 2014. [31] N. Singh, W. W. Fang, S. C. Rustagi, K. D. Budharaju, Selin H. G. Teo, S. Mohanraj, G. Q. Lo, N. Balasubramanian, and D.-L. Kwong. Observation of Metal-Layer Stress on Si Nanowires in Gate-All-Around High- 𝜅 /Metal-Gate Device Structures. IEEE Electron Device Letters, 28(7):558–561, 2007. [32] J. Ch Sublet, I. P. Bondarenko, G. Bonny, J. L. Conlin, M. R. Gilbert, L. R. Greenwood, P. J. Griffin, Petter Helgesson, Y. Iwamoto, V. A. Khryachkov, T. A. Khromyleva, A. Yu Konobeyev, N. Lazarev, L. Luneville, F. Mota, C. J. Ortiz, D. Rochman, S. P. Simakov, D. Simeone, Henrik Sjöstrand, D. Terentyev, and R. Vila. Neutron-Induced Damage Simulations: Beyond Defect Production Cross-Section, Displacement per Atom and Iron-Based Metrics. European Physical Journal Plus, 134(7):1–50, 2019. 71 [33] A. P. Thompson, L. P. Swiler, C. R. Trott, S. M. Foiles, and G. J. Tucker. Spectral Neighbor Analysis Method for Automated Generation of Quantum-Accurate Interatomic Potentials. Journal of Computational Physics, 285:316–330, 2015. 72 APPENDIX A. Multi-layer MARLOWE input The input file template for MARLOWE multilayer calculations is reproduced below. The working units in the &SURF block are defined by the lattice constant of the crystal defined in the first &XTAL block. They can be converted back into nanometers by multiplying by 0.1 × 5.4309. The quantities inside "«»" indicate variables which will be populated by the driver script. Marlowe nFET S i PKA i n nFET # # Generic finFET l a y o u t # # g a t e | S i nm | g a t e | S i nm | g a t e | S i nm | g a t e # # G a t e m a t e r i a l − Al , 5nm # # S i l i c o n EBND( 3 ) = 2 . 0 # R e c o m b i n a t i o n r a n g e s h o u l d be 4 . 7 5A # &MODL RDNML=T , F , T , F , F , F , TRAM= . FALSE . , SURFCE=3 , KLAY=7 , DELTA ( 1 ) = 1 . 0 E−10 , ICHAN= 1 0 0 0 0 0 , 1 0 0 0 0 , 1 0 0 , 5 0 / &XTAL LYR=7 , NEWS=1 , ALAT= 5 . 4 3 0 9 , CENTRE=6 , POLY=< > , RZ = 3 ∗ 0 . 0 , 3 ∗ 0 . 2 5 , DMAX= 0 . 7 5 , AXISA = 1 . 0 , 0 . 0 , 0 . 0 , AXISB = 0 . 0 , 1 . 0 , 0 . 0 / &XTAL LYR=6 , NEWS=1 , ALAT= 5 . 4 3 0 9 , CENTRE=6 , POLY=< > , RZ = 3 ∗ 0 . 0 , 3 ∗ 0 . 2 5 , 73 DMAX= 0 . 7 5 , AXISA = 1 . 0 , 0 . 0 , 0 . 0 , AXISB = 0 . 0 , 1 . 0 , 0 . 0 / &XTAL LYR=5 , NEWS=1 , ALAT= 5 . 4 3 0 9 , CENTRE=6 , POLY=< > , RZ = 3 ∗ 0 . 0 , 3 ∗ 0 . 2 5 , DMAX= 0 . 7 5 , AXISA = 1 . 0 , 0 . 0 , 0 . 0 , AXISB = 0 . 0 , 1 . 0 , 0 . 0 / &XTAL LYR=4 , NEWS=1 , ALAT= 5 . 4 3 0 9 , CENTRE=6 , POLY=< > , RZ = 3 ∗ 0 . 0 , 3 ∗ 0 . 2 5 , DMAX= 0 . 7 5 , AXISA = 1 . 0 , 0 . 0 , 0 . 0 , AXISB = 0 . 0 , 1 . 0 , 0 . 0 / &XTAL LYR=3 , NEWS=4 , ALAT= 5 . 4 3 0 9 , CENTRE=6 , POLY=< > , RZ = 3 ∗ 0 . 0 , 3 ∗ 0 . 2 5 , DMAX= 0 . 7 5 , AXISA = 1 . 0 , 0 . 0 , 0 . 0 , AXISB = 0 . 0 , 1 . 0 , 0 . 0 / &XTAL LYR=2 , NEWS=1 , ALAT= 5 . 4 3 0 9 , CENTRE=6 , POLY=< > , RZ = 3 ∗ 0 . 0 , 3 ∗ 0 . 2 5 , DMAX= 0 . 7 5 , AXISA = 1 . 0 , 0 . 0 , 0 . 0 , AXISB = 0 . 0 , 1 . 0 , 0 . 0 / &XTAL LYR=1 , NEWS=1 , ALAT= 5 . 4 3 0 9 , CENTRE=6 , POLY=< > , RZ = 3 ∗ 0 . 0 , 3 ∗ 0 . 2 5 , 74 DMAX= 0 . 7 5 , AXISA = 1 . 0 , 0 . 0 , 0 . 0 , AXISB = 0 . 0 , 1 . 0 , 0 . 0 / &ATOM NTYPE=2 , TYPE= ’ Al ’ , ’ Si ’ , Z= 13.0 , 14.0 , W= 2 6 . 9 8 2 , 2 8 . 0 8 5 5 , INEL=<>,< > , EQUIT=5.0 , < > , EBND= 3 . 0 , 2 . 0 , 2 . 0 , 3 . 0 , 2 . 0 , 2 . 0 , 4.63 ,2.0 ,2.0 , 4.63 ,2.0 ,2.0 , 3.0 ,2.0 ,2.0 , 3.0 ,2.0 ,2.0 , 4.63 ,2.0 ,2.0 , 4.63 ,2.0 ,2.0 , 3.0 ,2.0 ,2.0 , 3.0 ,2.0 ,2.0 , 4.63 ,2.0 ,2.0 , 4.63 ,2.0 ,2.0 , 3.0 ,2.0 ,2.0 , 3.0 ,2.0 ,2.0 , LOCK=14∗ −1 , ORDER( 1 , 1 ) = 1 . 0 , ORDER( 2 , 1 ) = 0 . 0 , ORDER( 1 , 2 ) = 1 . 0 , ORDER( 2 , 2 ) = 0 . 0 , ORDER( 1 , 3 ) = 0 . 0 , ORDER( 2 , 3 ) = 1 . 0 , ORDER( 1 , 4 ) = 0 . 0 , ORDER( 2 , 4 ) = 1 . 0 , ORDER( 1 , 5 ) = 1 . 0 , ORDER( 2 , 5 ) = 0 . 0 , ORDER( 1 , 6 ) = 1 . 0 , ORDER( 2 , 6 ) = 0 . 0 , ORDER( 1 , 7 ) = 0 . 0 , ORDER( 2 , 7 ) = 1 . 0 , ORDER( 1 , 8 ) = 0 . 0 , ORDER( 2 , 8 ) = 1 . 0 , ORDER( 1 , 9 ) = 1 . 0 , ORDER( 2 , 9 ) = 0 . 0 , ORDER( 1 , 1 0 ) = 1 . 0 , ORDER( 2 , 1 0 ) = 0 . 0 , ORDER( 1 , 1 1 ) = 0 . 0 , ORDER( 2 , 1 1 ) = 1 . 0 , ORDER( 1 , 1 2 ) = 0 . 0 , ORDER( 2 , 1 2 ) = 1 . 0 , ORDER( 1 , 1 3 ) = 1 . 0 , ORDER( 2 , 1 3 ) = 0 . 0 , ORDER( 1 , 1 4 ) = 1 . 0 / ORDER( 2 , 1 4 ) = 0 . 0 / &SURF ORIGIN = 0 . 0 , 0 . 0 , 0.0 , 0.0 ,0.0 , 9.20661 , 0.0 ,0.0 , 18.41322 , 75 0.0 ,0.0 , 27.61983 , 0.0 ,0.0 , 36.82644 , 0.0 ,0.0 , 46.03305 , 0.0 ,0.0 , 55.23966 , 0.0 ,0.0 , 64.44627 , DEPTH= 9 . 2 0 6 6 , 9 . 2 0 6 6 , 9 . 2 0 6 6 , 9 . 2 0 6 6 , 9 . 2 0 6 6 , 9 . 2 0 6 6 , 9 . 2 0 6 6 , EDGE= 9 2 . 0 6 6 , 1 8 . 4 1 3 , SIDES = 9 2 . 0 6 6 , 1 8 . 4 1 3 / &SIZE SLICE = 0 . 0 / &OUTP TRACE= 0 , 0 , 0 , Look =3 , INFORM=F , F , F , F , F , T / &PROJ LEAP=6 , LAIP =2 , RAIP=<>,<>,< > , REFIP=<>,<>,< > , MILLER = . FALSE . , THA=< > , PHI=< > , MAXRUN=1 , EKIP=< >/ # # End o f i n p u t # 76 APPENDIX B. Multi-layer TRIM input The input file template for TRIM multilayer calculations is reproduced below. The quantities inside "«»" indicate variables which will be populated by the driver script. ==> SRIM−2013.00 T h i s f i l e c o n t r o l s TRIM C a l c u l a t i o n s . I o n : Z1 , M1, E n e r g y ( keV ) , Angle , Number , Bragg Corr , AutoSave Number . <> <> <> <> 1 1 1 C a s c a d e s ( 1 =No ; 2 = F u l l ; 3 = S p u t t ;4 −5= I o n s ;6 −7= N e u t r o n s ) , Random Number Seed , R e m i n d e r s 2 <> 0 D i s k f i l e s ( 0 = no , 1 = y e s ) : Ranges , B a c k s c a t t , T r a n s m i t , S p u t t e r e d , C o l l i s i o n s , EXYZ . t x t 77 0 0 0 0 0 0 T a r g e t m a t e r i a l : Number o f E l e m e n t s & L a y e r s " S i ( 1 0 ) i n t o S i _ 1 +Al_1+ S i _ 2 +Al_2+ S i _ 3 +Al_3 " 6 6 P l o t T y p e ( 0 − 5 ) ; P l o t D e p t h s : Xmin , Xmax ( Ang . ) [ = 0 0 f o r Viewing F u l l T a r g e t ] 1 <> <> Target Elements : Z Mass ( amu ) Atom 1 = S i = 14 2 8 . 0 8 6 Atom 2 = Al = 13 2 6 . 9 8 2 Atom 3 = S i = 14 2 8 . 0 8 6 Atom 4 = Al = 13 2 6 . 9 8 2 Atom 5 = S i = 14 2 8 . 0 8 6 Atom 6 = Al = 13 2 6 . 9 8 2 Layer L a y e r Name / Width D e n s i t y S i ( 1 4 ) Al ( 1 3 ) S i ( 1 4 ) Al ( 1 3 ) S i ( 1 4 ) Al ( 1 3 ) Numb . Description ( Ang ) ( g / cm3 ) Stoich Stoich Stoich Stoich Stoich Stoich 1 " S i _ 1 " <> 2 . 3 2 1 2 1 0 0 0 0 0 2 " Al_1 " 50 2.702 0 1 0 0 0 0 3 " Si_2 " 50 2.3212 0 0 1 0 0 0 4 " Al_2 " 50 2.702 0 0 0 1 0 0 5 " Si_3 " 50 2.3212 0 0 0 0 1 0 6 " Al_3 " 50 2.702 0 0 0 0 0 1 0 T a r g e t l a y e r p h a s e s ( 0 = S o l i d , 1= Gas ) 0 0 0 0 0 0 T a r g e t Compound C o r r e c t i o n s ( Bragg ) 1 1 1 1 1 1 I n d i v i d u a l t a r g e t atom d i s p l a c e m e n t e n e r g i e s ( eV ) 15 25 15 25 15 25 I n d i v i d u a l t a r g e t atom l a t t i c e b i n d i n g e n e r g i e s ( eV ) 2 3 2 3 2 3 I n d i v i d u a l t a r g e t atom s u r f a c e b i n d i n g e n e r g i e s ( eV ) 4.7 3.36 4.7 3.36 4.7 3.36 S t o p p i n g Power V e r s i o n ( 1 = 2 0 1 1 , 0 =2 0 1 1 ) 0 78 DISTRIBUTION Email—Internal Name Org. Sandia Email Address Patrick Griffin 1300 pjgriff@sandia.gov Greg Frye 1330 gcfrye@sandia.gov Andrew Tonigan 1314 amtonig@sandia.gov Reza Arghavani 6833 rarghav@sandia.gov Matthew Sternat 1323 mrstern@sandia.gov Alex Salazar 1321 alesala@sandia.gov Jesse Jones 1321 jjones8@sandia.gov Marlene Bencomo 1321 mbencom@sandia.gov Technical Library 1911 sanddocs@sandia.gov 79 Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.